GLM 5.2 vs Claude Sonnet 4.6 vs MiniMax M3: Benchmarks Tested Side by Side

MiniMax M3 wins on cost. GLM 5.2 wins on speed and licence. Sonnet 4.6 wins on reasoning. Here's why, with the prices checked against the labs' own pages.

Three models. Three different labs. Three very different value propositions. GLM 5.2 is the open-weight coding powerhouse. Claude Sonnet 4.6 is the balanced mid-tier workhorse. MiniMax M3 is the budget multimodal challenger. Here is how they actually compare.

Worth clearing up before the numbers: GLM is a model family, not a different architecture from an LLM. GLM vs LLM: the difference explained covers what the name means today.

Test all three on your own workload.

BetterClaw routes GLM 5.2, Claude Sonnet 4.6, and MiniMax M3 through one agent config via BYOK. Switch models with a setting, not a rewrite. Free forever, not a trial. Start free → No credit card · 28+ providers · Zero markup

GLM 5.2 from Zhipu AI is the open-weight coding powerhouse with an MIT license and, when it launched, the highest Intelligence Index score of any open model. Claude Sonnet 4.6 from Anthropic is the balanced mid-tier workhorse with near-flagship intelligence at $3/$15 pricing. MiniMax M3 from MiniMax is the budget multimodal challenger that undercuts both on cost while claiming frontier coding performance.

All three launched within weeks of each other in early to mid 2026. All three target agent builders. All three have real strengths and real weaknesses that marketing pages do not mention.

This comparison covers verified benchmarks, actual API pricing, tool calling reliability, agent workflow suitability, and honest assessments of where each model falls short. No affiliate links. No cherry-picked numbers. The right choice depends entirely on what you are building and what you are willing to spend.

This page covers benchmarks and capability first, then the bill. If the bill is all you need, jump to what an agent actually costs per day, which runs the per-task, per-day and routing arithmetic. If you're choosing a model for an OpenClaw agent specifically, picking a model for an OpenClaw agent has the per-workload recommendations.

Verdict: MiniMax M3 wins on price at $0.30/$1.20 under 512K input. GLM 5.2 wins on speed, independent benchmark scores and licence freedom. Sonnet 4.6 wins on reasoning depth and tool-chain reliability. If cost decides it: M3. If you need to ship the weights: GLM 5.2. If mistakes cost money: Claude, though Sonnet 5 at $2/$10 has since replaced 4.6 on both price and benchmarks.

The Quick Answer

If you want the summary before the full breakdown:

• Pick GLM 5.2 when you need the strongest open-weight coding model, self-hosting rights under MIT, or the lowest token cost for coding-heavy agent workloads. $1.40/$4.40 per million tokens via API. Open weights on HuggingFace.
• Pick Claude Sonnet 4.6 (or Sonnet 5.5 at $2/$10, the newer and cheaper Sonnet) when you need the best all-around model at mid-tier pricing, computer use for GUI-based tasks, or the most mature tool calling implementation. $3/$15 per million tokens. Best balance of capability, safety, and developer experience.
• Pick MiniMax M3 when cost is the deciding factor, you need multimodal input (images and video), or you need long context at the cheapest price available. $0.30/$1.20 per million tokens on requests up to 512K input, doubling to $0.60/$2.40 above that. Cheapest of the three by a wide margin.
• Pick all three via BetterClaw when you want to route different tasks to different models based on cost and capability, or you are not sure which model fits your workload best and want to test them side by side.

What Each Model Actually Is

GLM 5.2

Developer: Zhipu AI, operating under the Z.ai brand. Beijing-based AI company spun out of Tsinghua University's Knowledge Engineering Group in 2019. Now publicly listed.

Released: June 13 to 16, 2026.

Architecture: 744 billion total parameters, approximately 40 billion active per token. Mixture-of-Experts design. Introduces IndexShare, which reuses a lightweight indexer across every four sparse-attention layers to reduce per-token compute by 2.9x at 1M context. Also ships an improved multi-token prediction (MTP) layer for speculative decoding that increases acceptance length by up to 20%.

Context window: 1 million tokens.

License: MIT. This is the most permissive license available. You can download the weights, run locally, fine-tune on proprietary data, deploy in commercial products, and redistribute without attribution requirements.

Reasoning modes: Two levels called High and Max (xhigh). High gives faster responses with reasonable reasoning depth. Max allocates maximum compute for the hardest problems.

Key benchmark numbers (third-party verified): Intelligence Index v4.1 score of 51 (highest open-weight model at the time of its release). Terminal-Bench 2.1: 81.0. SWE-bench Pro: 62.1. FrontierSWE: leading among open-weight models. BenchLM.ai ranked it #4 out of 124 models with 91/100. Design Arena Code Category: #1 globally for frontend generation from natural language.

Important note: Zhipu published zero benchmark numbers at launch. Every number above comes from third-party evaluations (Artificial Analysis, BenchLM.ai, Design Arena, community testing). This is unusual for a flagship release and worth noting, even though the third-party results have been consistently strong.

Claude Sonnet 4.6

Developer: Anthropic. San Francisco-based AI safety company.

Released: February 17, 2026.

Architecture: Not publicly disclosed. Closed-weight model available only through API (Anthropic, Amazon Bedrock, Google Vertex AI).

Context window: 1 million tokens, and since March 2026 the full window bills at standard rates. Anthropic's own pricing page is explicit that Claude 4.6 and later models "include the full 1M token context window at standard pricing" — a 900K-token request costs the same per token as a 9K one. There is no premium long-context tier to budget for. Prompt cache hits are 0.1x base input, so $0.30 per million tokens, with an optional 1-hour TTL at 2x write.

This is worth flagging because a lot of comparison pages (including an earlier version of this one) still describe Sonnet 4.6 as a 200K model with a paid 1M beta. That stopped being true in March.

Reasoning modes: Four adaptive thinking levels (low, medium, high, max). The model automatically adjusts reasoning depth to task difficulty, spending minimal overhead on simple tasks and full reasoning chains on complex problems.

Key benchmark numbers (Anthropic system card, independently validated): SWE-bench Verified: 79.6%. OSWorld-Verified: 72.5% (computer use). Terminal-Bench 2.0: 59.1%. ARC-AGI-2: 58.3% (a 4.3x improvement over Sonnet 4.5). GDPval-AA: 1633 Elo (best of all models for office productivity). Finance Agent: 63.3% (best-in-class). MCP-Atlas: 61.3%.

Developers preferred Sonnet 4.6 over the previous generation Sonnet 4.5 in 70% of head-to-head comparisons. They preferred it over the older flagship Opus 4.5 in 59% of comparisons. That is a mid-tier model beating the previous generation's premium flagship.

MiniMax M3

Developer: MiniMax. Shanghai-based AI lab founded in 2021. Listed on the Hong Kong Stock Exchange in January 2026.

Released: June 1, 2026.

Architecture: 428 billion total parameters, approximately 23 billion active per token. Mixture-of-Experts. Built on MiniMax Sparse Attention (MSA), which partitions the KV cache into blocks to cut per-token compute at long context to roughly 1/20th of the previous generation, with 9x+ faster prefill and 15x+ faster decoding.

Context window: 1 million tokens, with a guaranteed minimum of 512K and a billing tier change at exactly that line. Independent trackers list M3's usable context as 512K for that reason.

License: MiniMax Community License. Open-weight but with commercial conditions. Not MIT, and not a modified MIT either — MiniMax has said so directly. Commercial use requires displaying "Built with MiniMax M3" on your product or docs, a one-time notice to MiniMax, and prior written authorisation above roughly $20M in yearly revenue from the product. Read the terms before you ship.

Multimodal: Native text, image, and video input. The only model of these three that processes video.

Key benchmark numbers (company-reported, mostly unverified as of mid-June 2026): SWE-Bench Pro: 59.0%. Terminal-Bench 2.1: 66.0%. BrowseComp: 83.5%. SWE-fficiency: 34.8%. KernelBench Hard: 28.8%. MCP-Atlas: 74.2%. MiniMax claims scores surpassing GPT-5.5 and Gemini 3.1 Pro on coding and edging past Claude Opus 4.7 on autonomous browsing.

Important note: Most MiniMax M3 benchmark scores are from MiniMax's own testing infrastructure with their agent scaffolding. Independent verification is still pending as of mid-June 2026. Treat these numbers as indicative rather than confirmed. Artificial Analysis Intelligence Index v4.1 independently scored M3 at 44, which is above average but well below GLM 5.2's 51.

Pricing: The Numbers That Actually Matter

This is where the three models diverge most dramatically, and pricing drives most real-world model selection decisions. Every number below is the first-party rate from the lab's own pricing page, not an aggregator's cheapest reseller.

Get the M3 number right, because most pages have it wrong. Comparison articles routinely quote $0.60/$2.40 as M3's standard price with $0.30/$1.20 as a temporary launch promo, then warn you to budget on the higher figure. That is backwards. MiniMax's published standard rate is $0.30/$1.20 for any request with 512K or fewer input tokens. $0.60/$2.40 is not a rate you fall back to when a promo lapses; it is the tier you enter when a single request crosses 512K input tokens. MiniMax does label the displayed rates as a permanent 50% discount against struck-through list prices, but "permanent 50% off" has been the shipping price all year. Cost tables built on the $0.60 assumption are roughly 2x too high.

The 512K cliff is a step, not a slope. When one request crosses 512,000 input tokens, the entire call rebills at the doubled rate. Not the excess portion. A 600K-token call costs more than double a 500K one. Agent loops are exactly the workload that drifts across that line by accident, because every turn appends tool results and reasoning nobody pruned.

What a typical long-context agent cycle costs (1M input + 500K output):

• MiniMax M3, one 1M-token call: $0.60 + $1.20 = $1.80 (over the cliff, doubled tier)
• MiniMax M3, same work split into two sub-512K calls: $0.30 + $0.60 = $0.90
• GLM-5.3-Flash: $0.15 + $0.25 = $0.40
• GLM 5.2 or GLM 5.3 (same price): $1.40 + $2.20 = $3.60
• Claude Sonnet 5: $2.00 + $5.00 = $7.00
• Claude Sonnet 4.6: $3.00 + $7.50 = $10.50
• Claude Opus 5.5: $4.00 + $10.00 = $14.00

Scaled to 100 agent runs per day for a month (3,000 runs):

• GLM-5.3-Flash: ~$1,200/month
• MiniMax M3, pruned under 512K: ~$2,700/month
• MiniMax M3, one call over the cliff: ~$5,400/month
• GLM 5.2 or 5.3: ~$10,800/month
• Claude Sonnet 5: ~$21,000/month
• Claude Sonnet 4.6: ~$31,500/month
• Claude Opus 5.5: ~$42,000/month

Note what that list actually shows: pruning your context below 512K saves more on M3 than picking M3 over GLM 5.2 saves you in the first place.

The gap is enormous at scale. But pricing without quality context tells you nothing. A model that costs half as much but needs twice as many retries to get a correct answer is not actually cheaper. Keep reading.

Where cost comparison gets nuanced: Sonnet 4.6's prompt caching ($0.30 per million tokens for cache hits, 90% cheaper than fresh input) dramatically changes the economics for workflows with repeated system prompts or shared context. If your agent reuses a long system prompt across many queries, Sonnet 4.6's effective per-query cost drops substantially. GLM 5.2's cache pricing ($0.26/M) is similar but less documented, and M3's cache reads at $0.06/M are the cheapest of the three outright. The cost section below runs the caching arithmetic on a real task.

Benchmark Comparison

Here are the benchmarks that matter most for agent builders, with verified numbers where available and clear notes where numbers are self-reported.

Reading the table honestly: Sonnet 4.6 has the most comprehensive and independently validated benchmark profile of the three. GLM 5.2 has strong third-party numbers on coding benchmarks. MiniMax M3 has impressive self-reported numbers that still have thinner independent confirmation than either of the others.

Two traps in that table are worth naming, because they are the two most common ways model comparisons mislead people.

SWE-bench Pro and SWE-bench Verified are not the same benchmark. You cannot put M3's 59.0% on SWE-bench Pro next to a Claude model's 80-something percent on SWE-bench Verified and conclude anything. The test sets, evaluation criteria and difficulty distributions differ. Any page that lines those two columns up side by side and declares a winner is doing arithmetic on incompatible units.

Intelligence Index scores are version-scoped. Artificial Analysis reweights its suite periodically, and scores from different versions are not comparable. M3 was reported at 55 on the index in early June, then 44 once v4.1 landed on 15 June and shifted the weighting toward newer agentic workloads, then lower again on v4.3. The GLM 5.2 (51) vs M3 (44) comparison in this post is v4.1 against v4.1, which is the only way that comparison means anything. If you see those numbers quoted against a score from a different index version, it is not a like-for-like gap.

Also remember who ran the test. A lab reporting its own benchmark on its own infrastructure with its own agent scaffolding is publishing a ceiling, not a floor. Kili Technology, which runs production AI evaluations, measured a 37% average gap in 2026 between lab benchmark scores and real-world deployment performance for enterprise agentic systems. That gap is a property of the problem, not a flaw in any one model: benchmarks measure capability, production measures reliability.

Tool Calling and Agent Suitability

For anyone building agents, these are the details that benchmarks do not fully capture.

GLM 5.2 Tool Calling

GLM 5.2 supports native function calling, structured JSON output, and extended reasoning with two effort levels. The 1M context window means you can feed an entire codebase into the prompt and maintain conversation history without chunking.

Strengths: Sustains quality over very long coding sessions. The model can chain hundreds of tool calls in coding agent workflows. MIT license means you can deploy it on your own infrastructure with complete control. Design Arena ranked it #1 globally for frontend code generation from natural language, which speaks to practical coding utility beyond benchmark scores.

Weaknesses: Text-only. No image or video input whatsoever. The model tends to be verbose (generating roughly 27% more tokens than average on Intelligence Index evaluation), which can inflate costs on output-priced APIs. The ecosystem around GLM models is smaller than Claude's or OpenAI's, so fewer pre-built integrations exist. Independent benchmark coverage is still catching up since the model is less than two weeks old as of this writing.

Claude Sonnet 4.6 Tool Calling

Sonnet 4.6 has the most mature and battle-tested tool calling implementation of the three. Anthropic has been iterating on tool use since October 2024, and the infrastructure shows.

Strengths: Interleaved tool calls during extended thinking (the model can use tools mid-reasoning without breaking its chain of thought). Strict JSON mode validates outputs server-side against declared schemas. 64% reduction in tool-call latency versus the previous Sonnet 4.5. Best-in-class computer use at 72.5% OSWorld, meaning the model can interact with GUIs, click buttons, fill forms, and navigate web interfaces. Strong prompt injection resistance, performing on par with Opus 4.6. Adaptive thinking automatically adjusts reasoning depth to task difficulty without manual configuration.

Weaknesses: Most expensive of the three at $3/$15 per million tokens, and roughly 10x M3's sub-512K rate. Closed-weight model with no self-hosting option and no way to inspect what you are running. Constitutional AI safety guardrails can occasionally result in refusals on edge-case tasks that other models handle without friction. And since Sonnet 5 shipped at $2/$10, Sonnet 4.6 is now the more expensive Sonnet, which is an odd place for it to sit.

MiniMax M3 Tool Calling

M3 supports function calling and demonstrated autonomous operation in MiniMax's internal showcases: a 12-hour ICLR paper reproduction with 18 commits and 23 experimental figures, and a 24-hour kernel optimization run with 147 benchmark submissions.

Strengths: Native multimodal input (text, image, video) gives it capabilities the other two simply do not have. Long context at $0.30/$1.20 under 512K is the most affordable long-context inference of these three by a wide margin, and cache reads at $0.06/M are the cheapest line item on this page. MiniMax Sparse Attention makes long-context work genuinely cheap. The model supports a thinking on/off toggle per request.

Weaknesses: Community tooling, tutorials, and integration support are still maturing compared to Claude's extensive ecosystem — eight tracked providers against GLM 5.2's seventeen. Benchmark scores are mostly company-reported and thinly verified by independent labs. The commercial license requires review before deployment (not MIT like GLM 5.2), and the attribution plus revenue-threshold clauses are real obligations, not boilerplate. The 512K billing cliff punishes exactly the long-horizon agent loops the model is marketed for. MiniMax is headquartered in Shanghai, which raises data sovereignty considerations under China's 2017 National Intelligence Law for teams processing sensitive data through the MiniMax API.

Speed, and Why Your Provider Matters More Than Your Model

GLM 5.2 wins throughput, and not narrowly.

Head to head on Artificial Analysis, GLM 5.2 in max mode generates around 150 output tokens per second against roughly 76 for M3. Time to first token is close, 1.39 seconds versus 1.46, so the gap is throughput rather than responsiveness. Median across all providers puts GLM 5.2 near 213 tokens per second, which is unusually quick for an open-weight model of that size. Sonnet 4.6 sits around 50-80 tokens per second depending on provider and load.

Read the provider column before you trust any single figure, though. GLM 5.2 is served by seventeen tracked providers with a 1,064% spread between fastest and slowest. M3 has eight, with a 296% spread. Your provider choice moves latency more than your model choice does, which is an argument for a platform that lets you swap providers without touching code.

M3 has an architectural advantage these numbers hide. MiniMax Sparse Attention keeps per-token compute at 1M context to roughly a twentieth of the previous generation, with decoding several times faster at the full window. At ordinary context lengths GLM is faster. At extreme ones, M3 degrades better.

Head-to-Head on Real Tasks

Task 1: Multi-File Code Refactoring

GLM 5.2 wins this category. The combination of 1M context, the strongest open-weight SWE-bench Pro score (62.1%), and sustained quality over long coding sessions makes it the top pick for repository-level work. It can hold a meaningful portion of a large codebase in context and produce consistent edits across multiple files without losing track of earlier changes.

Sonnet 4.6 is very close. 79.6% on SWE-bench Verified is near-flagship performance. For most day-to-day coding tasks, the gap between GLM 5.2 and Sonnet 4.6 is not noticeable in practice. Sonnet 4.6 tends to produce cleaner, more readable code with better variable naming and documentation, and its 1M window at flat pricing covers any real-world refactoring scope.

M3 is solid but needs time. 59% SWE-bench Pro is strong on paper, but without independent verification the actual gap to the other two is unclear. The BrowseComp score suggests strong autonomous capability, but coding refactoring and web browsing test different skills.

Task 2: Tool Use and Agent Workflows

Sonnet 4.6 wins. Most mature implementation, best latency numbers, and the only model with production-proven computer use. If your agent needs to interact with web interfaces, fill forms, navigate applications, or handle multi-step tool sequences with error recovery, Sonnet 4.6 is the clear choice.

GLM 5.2 is strong for coding-specific tool use. File operations, terminal commands, API calls, and test execution work well. The model handles the tool-call-execute-evaluate loop reliably for software engineering tasks.

M3 shows promise on agent benchmarks. The MCP-Atlas and BrowseComp scores suggest strong potential, but the production track record is too thin to recommend for mission-critical agent deployments today.

Task 3: Long Document Processing

All three now offer 1M tokens, and the differences are in the fine print. GLM 5.2 publishes a flat 1M input window with a separate output ceiling of about 131K tokens, which is roughly five times what GLM 5.1 allowed and matters if your agent writes long documents in one pass. Sonnet 4.6 bills its full 1M window at standard rates with no premium tier. M3's 1M is a ceiling over a 512K guaranteed minimum, that maximum counts input and output combined, and crossing 512K doubles the whole request.

So the ranking flips depending on where in the window you live. Under 512K, M3 is the cheapest long-context option available by a factor of four or more. Above it, M3's doubled tier ($0.60/$2.40) still beats GLM 5.2's flat $1.40/$4.40, but the margin narrows and the latency variance widens. Sonnet 4.6 costs the most per token at any length and is the only one of the three whose long-context quality has been independently validated at scale.

The practical move regardless of model: prune between turns. Agent loops fill context without asking permission, and by turn ten or fifteen you are dragging history no single step needed.

Task 4: Multimodal Tasks (Images, Video, Screenshots)

M3 wins by default. It is the only model of the three that accepts image and video input natively. GLM 5.2 is text-only. Sonnet 4.6 accepts images but not video. If your agent needs to understand screenshots, analyze UI designs, interpret charts, or process video frames, M3 is the only option among these three.

Task 5: Office Productivity and Business Tasks

Sonnet 4.6 wins decisively. Best of all models at 1633 Elo on GDPval-AA for office productivity. 63.3% on Finance Agent (also best-in-class). If your agent handles business documents, spreadsheets, email drafting, meeting summaries, or financial analysis, Sonnet 4.6 outperforms both alternatives on these specific tasks.

Open Weights vs Closed: Why It Matters for Agent Builders

This is not an academic distinction. It determines what you can build, where you can deploy, and who controls your infrastructure.

GLM 5.2 (MIT License, Open Weights): Download the weights. Run locally. Fine-tune on your data. Deploy on your infrastructure. Build commercial products. Redistribute modified versions. No attribution required. The practical constraint is hardware: the full model at BF16 is 1.51TB. At 2-bit quantization via Unsloth GGUF, it compresses to roughly 239GB, fitting on a Mac with 256GB unified memory or a workstation with 2+ A100 GPUs.

MiniMax M3 (MiniMax Community License, Open Weights): Open-weight but with commercial conditions. Self-hosting is possible but requires 75 to 150GB of memory at Q4 quantization (Mac Studio 192GB or 2+ A100s). Ollama offers M3 as a cloud-hosted model (minimax-m3:cloud) for zero-setup access. Review the license terms before commercial deployment: attribution is mandatory, and past roughly $20M in annual product revenue you need written authorisation.

The predecessor generation is worth knowing about too. GLM 5.1, the April model that put Z.ai on the map, is MIT with the same $1.40/$4.40 first-party pricing, a 200K context window, and 58.4% on SWE-bench Pro. Its distinctive claim was endurance: Z.ai demonstrated it running autonomously for eight hours on a single coding task, 655 iterations without human intervention. That is a different capability from a big context window, and it is still the reason some teams run GLM for overnight refactors and multi-hour research jobs rather than the model with the larger window. At 754B total and 42B active it is the heaviest of these to self-host; none of these models is practical on consumer hardware, and if you need something that runs on a laptop you are looking at a 30B-class model with 3B active parameters, not any of these.

Claude Sonnet 4.6 (Closed): No weights available. API-only through Anthropic, Amazon Bedrock, or Google Vertex AI. Cannot self-host, fine-tune, or inspect. What you get in exchange: the most thoroughly tested safety layer, the best developer documentation, the most extensive integration ecosystem, and consistent behavior across deployments.

For teams where cost at high volume and infrastructure control matter most, GLM 5.2's MIT license is a genuine competitive advantage. For teams where reliability, safety, and time-to-production matter most, Sonnet 4.6's closed ecosystem is not a limitation. It is the product.

Self-hosting vs the API: where the math flips

GLM 5.3 shares GLM 5.2's architecture, so it's multi-GPU territory at full precision and roughly half that with FP8 quantisation. At low volume the API at $1.40/$4.40 is cheaper than owning the hardware. The math flips somewhere around 500M to 1B tokens a month.

GLM-5.3-Flash moves this more than anything else. At 320B total and 18B active it's a fraction of the full GLM footprint, and it's MIT. If self-hosting was out of reach for your team before, Flash is the first model in this comparison where it's genuinely approachable.

The bigger hidden cost is operational. Self-hosting means running vLLM or SGLang, handling model updates, managing quantisation tradeoffs and debugging your own inference stack. It's a bet that your engineering time is cheaper than the API margin, and the hidden infrastructure costs add up quickly. For most teams, the API is the right answer.

Data sovereignty applies to both Chinese labs

MiniMax (Shanghai) and Z.ai (Beijing) both process data on Chinese-operated infrastructure by default when you call their own APIs. That's a compliance question rather than a technical one. For healthcare, finance, government or any other regulated workload, it may rule both labs' APIs out regardless of cost.

There are two ways around it: use a Western hosting provider for the open weights instead of the lab's API, or self-host. Both families' weights are downloadable, but the licences differ sharply, so read yours before you budget the GPUs.

The Complete Comparison Table

What changed in October 2026

The June head-to-head above is dated history and stays as it was. The price list around it has moved again. Every price below was checked on 13 October 2026 on the provider's own pricing page.

• Claude Sonnet 5.5 is now the newest Sonnet. $2/$10 per million tokens, cache hits $0.10, batch $1/$5, 1M context. Same headline price as Sonnet 5, with cache hits at half Sonnet 5's $0.20. Sonnet 5 stays on sale at $2/$10 as a legacy model.
• Sonnet 4.6 is still on sale at $3/$15 (cache hits $0.30, batch $1.50/$7.50), but it is no longer the newest or the cheapest Sonnet. Sonnet 5 and Sonnet 5.5 both cost a third less on input and output.
• Claude Haiku 5.5 (7 October). $0.10/$0.50 with $0.01 cache hits for prompts up to 100K tokens. A prompt over 100K bills the whole request at $0.50/$2.50. 1M context. Anthropic says the newer tokenizer produces roughly 30% more tokens for the same text than older models, so the effective saving over Haiku 4.5 ($1/$5) is a little smaller than the rate card suggests.
• Claude Opus 5.5 is unchanged at $4/$20 (cache hits $0.20, batch $2/$10) and is still the current Opus.
• GPT-6.1 Sol lists at $2/$10 with $0.10 cached input and a 1.05M context window, the same headline price as Sonnet 5.5. Prompts over 272K input tokens are billed at 2x input and 1.5x output for the whole request.
• The cheap tier is unchanged at the top and has a new peak/off-peak entrant. GLM-5.3 is still $1.40/$4.40 ($0.26 cached) and GLM-5.3 Flash $0.15/$0.50 ($0.03 cached). DeepSeek V4.1 Flash is $0.30/$1.20 at peak (01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays) and half that, $0.15/$0.60, the rest of the time. 1M context.

None of the new models has been through the head-to-head tests on this page, so there are no benchmark claims for them here. The new rows are in the full rate card and the daily cost table below.

Earlier in 2026: GLM 5.3, Sonnet 5 and Opus 5.5

All three labs shipped again over the summer. These are the releases that first moved the verdict. Prices below are the labs' own rates, first checked on 28 September 2026 and re-checked on 13 October 2026.

GLM 5.3 (Z.ai, 18 August 2026). Built on exactly the same base weights as GLM 5.2, with every improvement coming from post-training. 1M context, text-only, 128K max output, reasoning always on with low, high and max effort levels. Priced identically to 5.1 and 5.2 at $1.40/$4.40, cached input $0.26. Z.ai has held that line across three generations, which is why quoting GLM at $0.98 or $1.00 from a reseller's page misstates what the lab charges. Weights landed on Hugging Face on 25 August.

Z.ai's model card numbers against GLM 5.2 (all lab-reported, run in Claude Code as the harness): Terminal-Bench 2.1 88.2 vs 81.0, Terminal-Bench 3.0 28.3 vs 4.6, DeepSWE 66.9 vs 46.2, Toolathlon Verified 73.0 vs 59.9, CyberGym 84.5 vs 77.2. Z.ai did not publish a SWE-bench Verified or SWE-bench Pro score for 5.3, so there's no like-for-like SWE-bench comparison with Sonnet 5.

The licence changed, and most comparison pages missed it. GLM 5.1 and 5.2 are MIT. GLM 5.3 ships under a custom "GLM-5.3 License". It reads like MIT, with one added clause: any company running a model-as-a-service business with more than $10 billion in revenue over 12 months has to pass a Z.ai security review before using it commercially. For almost every team reading this, it's still effectively permissive. But "MIT" is no longer accurate for the flagship. GLM-5.3-Flash is still MIT.

GLM-5.3-Flash (Z.ai, 26 August 2026). A 320B-total, 18B-active MoE with a 1M context window, native image and video input, and MIT-licensed open weights. It's the first natively multimodal model in the GLM-5 series, which takes away M3's claim to be the only multimodal open-weight option. Z.ai lists it at $0.15/$0.50 per million tokens with cached input at $0.03. That's half M3's sub-512K rate, and there's no context cliff. A faster FlashX variant costs $0.37/$1.25. It's a small model at a price nothing else on this page matches.

Claude Sonnet 5 (Anthropic, GA 30 June 2026). $2/$10 per million tokens, cache hits at $0.20, batch at $1/$5, and a 1M context window at standard pricing. Anthropic launched $2/$10 as an introductory rate and scheduled a rise to $3/$15 for 1 September. It then cancelled the rise: $2/$10 is now the standard price. Its system card reports 85.2% on SWE-bench Verified, 63.2% on SWE-bench Pro and 80.4% on Terminal-Bench 2.1. It's cheaper and stronger than Sonnet 4.6 on every one of those, which makes Sonnet 4.6 hard to recommend for a new build. Sonnet 4.6 is still available at $3/$15 if you have a pinned workload.

Claude Opus 5.5 (Anthropic, 22 September 2026). $4/$20, cache hits at $0.20, batch at $2/$10, and a research-preview Fast mode at $8/$40. It replaces Opus 5 ($5/$25, released 24 July) as the current Opus model. Anthropic's launch table has it at 66.4% on Terminal-Bench 4.0 against Opus 5's 52.3%, and 81.8% on OSWorld 2.0 against 74.0%. Those are newer benchmark versions than anything else on this page, so don't line them up against the Terminal-Bench 2.1 numbers above.

GLM 5.3 vs Sonnet 5, head to head. On price, GLM 5.3 is 1.4x cheaper on input and 2.3x cheaper on output. The June gap against Sonnet 4.6 was 2.1x and 3.4x. On Terminal-Bench 2.1, Z.ai reports 88.2 and Anthropic reports 80.4 for Sonnet 5. Each lab ran its own harness, so treat that as a direction, not a margin. Sonnet 5 still takes images, and GLM 5.3 still doesn't.

One caution on the Claude 4.7-and-later models, which include Sonnet 5 and Opus 5.5: they use a newer tokenizer that produces roughly 30% more tokens for the same text. A per-million-token price cut doesn't translate one-for-one into a bill cut if the token count moves with it.

The short version if you're starting fresh today: GLM-5.3-Flash is the new cost floor. Sonnet 5 replaces Sonnet 4.6 as the sensible Anthropic default. GLM 5.3 is the coding value pick if you can live with text-only. Opus 5.5 is for the hard cases. M3 keeps its niche in cheap multimodal long context under 512K.

What an Agent Actually Costs Per Day

Per-million-token rates are hard to reason about. Here's the same set of models turned into per-task and per-day bills, with the current line-up included alongside the June three.

The full rate card

First-party rates per million tokens, from each lab's own pricing page, checked 13 October 2026.

Anthropic's Batch API halves the Claude rows. Sonnet 5 and Sonnet 5.5 drop to $1/$5, Haiku 5.5 to $0.05/$0.25 (prompts up to 100K), Sonnet 4.6 to $1.50/$7.50 and Opus 5.5 to $2/$10. For anything that isn't latency-sensitive, like nightly enrichment, backfills, bulk classification or scheduled agent jobs, that's a 50% cut you don't have to negotiate. Neither MiniMax nor Z.ai offers a batch tier. OpenAI lists Batch at 50% off for GPT-6.1 Sol.

Reseller rates aren't the lab's rate. Pages that quote GLM at $0.98/$3.08 or $1.00/$3.20 are quoting an aggregator, or (in the $1/$3.20 case) the older GLM-5. The gap matters. Budget a 500-task day (profile below) at $0.98/$3.08 and you get $2.70. The first-party bill is $3.86, 43% more than you planned for.

Cost per task: 500K input, 100K output

This is a context-heavy agent task: a long document or a fat accumulated conversation, plus a substantial written answer. 500K input keeps M3 just under its cliff. Haiku 4.5 is left out because its 200K window can't take the input.

With prompt caching, say 80% of that 500K input is a repeated system prompt and shared context. Sonnet 4.6 then pays $3/M on 100K fresh input ($0.30), $0.30/M on 400K cached ($0.12), and full price on output ($1.50), for $1.92. Sonnet 5 lands at $1.28. Caching is a real lever, but output is never cached, and on the Claude models output is where most of the money goes. A page claiming caching takes Sonnet 4.6 down to $0.90 on this task has forgotten the output line.

Push the same task over 512K input and M3 doubles to about $0.54 while nothing else moves.

Daily cost at 100, 500 and 2,000 tasks

Now the numbers people actually budget against. Token profile: 3,000 input tokens per task (system prompt, user message, tool definitions) and 800 output tokens. No caching, so treat these as ceilings. Monthly is 30 days.

DeepSeek V4.1 Flash off-peak halves its row. The Claude 4.7-and-later rows (Haiku 5.5, Sonnet 5, Sonnet 5.5, Opus 5.5) assume the same token counts as the others; Anthropic's newer tokenizer produces roughly 30% more tokens for the same text, so budget a little above these.

At 100 tasks a day (a personal agent), the whole spread is about $82 a month. Pick on quality, not cost.

At 500 tasks a day (a business agent working support tickets, lead qualification or CRM updates), the gap starts to mean something. Sonnet 5 costs $182 a month more than M3, which is $2,184 a year per agent.

At 2,000 tasks a day (production), all-Opus costs about $1,570 a month more than all-M3. This is where routing stops being a nice idea and becomes the architecture.

The failure-cost line nobody puts in the table

Per-token price is only half the comparison. The other half is what a failed task costs you.

Claude Sonnet's measured tool-call hallucination rate is about 3%, against roughly 10% for the budget tier. Treat the budget-tier figure as a rough estimate, since it hasn't been independently verified. On 500 daily tasks with five tool calls each, that works out to about 75 failures a day on Sonnet and about 250 on M3. If each extra failure takes five minutes of a $30/hour reviewer's time, the 175 extra failures cost around $437 a day in review labour. Choosing M3 over Sonnet 5 saves about $6 a day. Sonnet pays for its premium once a wrong answer costs you more than about four cents.

Two more ways the headline rate misleads you:

• Retries are 100% surcharges. A model that's 4x cheaper per token but needs a second attempt on a third of tasks has spent most of its advantage. GLM 5.2 scores 51 on Intelligence Index v4.1 against M3's 44 on the same version, so it should need fewer second attempts.
• Verbosity multiplies output price. GLM 5.2 generates roughly 27% more tokens than average, and that extra output costs more at $4.40/M than the same verbosity would at $1.20/M. GLM's real cost sits further above its headline rate than M3's does.

The routing math, worked

Route 65% of tasks (classification, extraction, formatting) to M3, 25% (coding, drafting, analysis) to GLM 5.2, and 10% (complex reasoning, customer-facing) to Sonnet 5. At 2,000 tasks a day, that's 1,300 / 500 / 200.

• M3, 1,300 tasks: 3.9M input at $0.30 = $1.17, plus 1.04M output at $1.20 = $1.25. $2.42/day
• GLM 5.2, 500 tasks: 1.5M input at $1.40 = $2.10, plus 0.4M output at $4.40 = $1.76. $3.86/day
• Sonnet 5, 200 tasks: 0.6M input at $2.00 = $1.20, plus 0.16M output at $10.00 = $1.60. $2.80/day

Routed total: $9.08/day, about $272/month. That's 68% less than putting everything on Sonnet 5 ($840) and 78% less than all-Sonnet 4.6 ($1,260), and you still get Claude-grade output on the 200 tasks a day that need it. It's 2.4x the all-M3 bill of $112, which is what it costs to avoid most of those extra tool-call failures.

Swap Sonnet 5 for Opus 5.5 on that top 10% and the routed total moves to $11.88/day (about $356/month). That's still well under half of all-Sonnet-5, with flagship reasoning on the hard tasks.

Which One Should You Use?

Use GLM 5.2 if:

• Cost per token is a primary concern and you run high-volume coding agent workloads
• You need MIT-licensed open weights for self-hosting, fine-tuning, or compliance
• Your workload is primarily coding and text processing (no multimodal needs)
• You want the strongest open-weight model available for software engineering tasks
• Infrastructure independence matters (no single API provider dependency)

Use Claude Sonnet 4.6 (or Sonnet 5.5 at $2/$10, the newer and cheaper Sonnet) if:

• You need the best overall model balancing coding, tool use, and general tasks
• Computer use (interacting with GUIs, filling forms, navigating web apps) is part of your workflow
• You want the most mature, battle-tested tool calling with lowest latency
• Safety, prompt injection resistance, and reliable behavior matter for your deployment
• You are already in the Anthropic ecosystem (Claude Code, Bedrock, Cowork)
• Office productivity and business document tasks are core to your use case

Use MiniMax M3 if:

• Budget is the deciding factor and you need frontier-adjacent performance at a fraction of the cost
• Your agent needs to understand images or video (screenshots, charts, visual content, video frames)
• You need long context at the cheapest price available and can keep single requests under 512K
• You are comfortable with a model that has less independent benchmark verification than the other two
• You have evaluated the data sovereignty implications and the Community License terms for your specific use case

If you want a closer two-way read, we also break down GLM 5.2 vs Sonnet 4.6 and, for the Anthropic side specifically, MiniMax M3 vs Claude Sonnet 4.6 — which is where the Opus-tier routing math lives.

Picking a Model for an OpenClaw Agent

If the agent you're building runs on OpenClaw, the choice changes shape.

Why agents change the math

A chat app answers one question per request. An OpenClaw agent loops. It reads a tool output, decides what to do next, calls another tool, reads that output and decides again. One user request can trigger 20 or 30 model calls internally.

That has three consequences:

• Reliability compounds. A model that's 10% more reliable cuts retry loops on every step of the chain, not just once.
• Per-token savings compound too. A model that's 5x cheaper per token ends up far cheaper per completed task, because it's billed on every loop iteration.
• Context grows on its own. Every turn appends tool results and reasoning nobody pruned. On M3, that's how a cheap agent drifts across 512K and silently doubles its own bill.

For a chat app, pick the smartest model you can afford. For an agent, pick the one that finishes the most tasks per dollar.

Recommendations by workload

High-volume, simple tasks (lead scoring, email classification, data extraction, triage, log summarisation, content moderation): GLM-5.3-Flash, or MiniMax M3 when the agent needs video input or routinely works near the top of a million-token window. At these rates, a classification agent running 10,000 items a day costs single-digit dollars a month. Flash is MIT-licensed and has no context cliff.

Multi-step workflows (email triage with CRM lookup, support ticket resolution, research summaries with tool calls): test the cheap model with Claude as the fallback. Run M3 or Flash for one to two weeks. Measure tool-call success rate, output quality and cost per resolved task, not per token. If the success rate lands within 5% of Claude's, the savings justify the switch. If the gap is wider, keep the cheap model on the simple steps and send the reasoning-heavy step to Claude.

Customer-facing or high-stakes work (support conversations, financial processing, compliance-sensitive operations): Claude Sonnet 5 with prompt caching, with Opus 5.5 as the escalation for the hardest reasoning. Most of the old reasons to pay for Opus now apply only to a narrow set of deep-reasoning and long-horizon tasks. Our Sonnet vs Opus for OpenClaw breakdown covers where that line sits.

Agentic coding and long-running engineering (code review, autonomous code generation, repository-wide refactors): GLM 5.3 via the Z.ai API. You get the best price-to-capability ratio on coding, a 1M window, open weights if you ever need them, and a lineage with an eight-hour, 655-iteration unattended run behind it. In April, GLM 5.1 scored 58.4 on SWE-bench Pro against Claude Opus 4.6's 57.3. That put an open-weight, MIT-licensed model ahead of Anthropic's flagship of the time on a serious agentic coding benchmark.

An agent doing all of the above: route by task. Flash for triage, GLM 5.3 for long coding sessions, Sonnet 5 for anything the user sees. See our model routing guide for the setup.

Still on Sonnet 4.6, GLM 5.1 or MiniMax M2.7?

All three have direct successors that are cheaper, stronger or both, and none of the moves require rewriting the agent:

• Sonnet 4.6 → Sonnet 5. It's a third cheaper ($2/$10 vs $3/$15) and scores higher on SWE-bench Verified (85.2% vs 79.6%). If a production agent is pinned to 4.6, do this before you evaluate anything else on this page.
• GLM 5.1 → GLM 5.3. Same $1.40/$4.40 price, with a 1M window instead of 200K. Check the licence first if you self-host: 5.1 is MIT, 5.3 is the GLM-5.3 License.
• MiniMax M2.7 → M3. M2.7 was $0.30/$1.20 with a 200K window, 56.2% on SWE-Pro and 57.0% on Terminal-Bench 2, under a non-commercial licence. M3 keeps the headline rate under 512K input and adds a 1M ceiling, native image and video input, and 59.0% on SWE-bench Pro.

Wiring the models in

GLM exposes an Anthropic-compatible endpoint alongside its OpenAI-shaped one, and Z.ai has called out OpenClaw compatibility in its release documentation. MiniMax is OpenAI-compatible. Most swaps come down to a base URL and a key. For a multi-model setup, pick a model for each category of task, configure keys for each provider, set the routing rules, and test the fallback path for when one provider is down. On self-hosted OpenClaw you manage the provider SDKs, routing logic and credential storage yourself.

If you'd rather not pay for tokens while you prototype, our guide to the best free model for OpenClaw covers the zero-cost options. For a wider field than these three, see how models compare for OpenClaw workloads.

Access All Three Through BetterClaw

BetterClaw supports BYOK across 28+ model providers. Connect to GLM 5.2 through OpenRouter or the Z.ai API. Access Claude Sonnet 4.6 through Anthropic directly. Use MiniMax M3 through OpenRouter or the MiniMax API. One agent configuration, multiple model backends, zero infrastructure to manage.

Test each model on your actual workload. See which one produces the best results for your specific use case. Switch between them by changing a setting, not rewriting your agent. If you are routing tasks across models to control spend, our model routing guide walks through the setup.

Get started with BetterClaw for free. Free plan includes 1 agent and 100 credits a month. Free forever, no credit card required.

Frequently Asked Questions

Is GLM 5.2 better than Claude Sonnet 4.6 for coding?

On pure coding benchmarks, GLM 5.2 scores higher. Terminal-Bench 2.1: 81.0% vs 59.1%. SWE-bench Pro: 62.1% vs an estimated 55%. On SWE-bench Verified (real GitHub issue resolution), both models land near 80%, close enough that practical differences depend on your specific codebase and task type. Sonnet 4.6 has the edge on tasks requiring computer use, GUI interaction, or combined coding plus business reasoning. GLM 5.2 wins on raw coding throughput, especially at scale where $1.40/$4.40 against $3/$15 is a 2.1x input and 3.4x output cost advantage. Note that Claude Sonnet 5 has since shipped at $2/$10 with 63.2% on SWE-bench Pro and 80.4% on Terminal-Bench 2.1, which narrows both the price and the benchmark gap considerably.

How much does MiniMax M3 cost compared to Claude Sonnet 4.6?

MiniMax's published standard rate is $0.30 per million input tokens and $1.20 per million output tokens for requests up to 512K input, which is 10x cheaper than Sonnet 4.6 on input and 12.5x cheaper on output. Above 512K input the whole request rebills at $0.60/$2.40, which is still 5x and 6.25x cheaper. Against the newer Claude Sonnet 5 at $2/$10, the sub-512K gap is about 6.7x on input and 8.3x on output. Be careful with pages that describe $0.60/$2.40 as M3's standard price and $0.30/$1.20 as a temporary promo — that has the tiers backwards, and it makes their cost tables roughly twice as expensive as reality.

Can I run GLM 5.2 locally?

Yes, but it requires serious hardware. The full BF16 checkpoint is 1.51TB. At 2-bit quantization (Unsloth Dynamic GGUF), it compresses to approximately 239GB and needs roughly 245GB+ of available memory. This fits on a Mac with 256GB unified memory or a workstation with 2+ NVIDIA A100 GPUs. Ollama lists glm-5.2:cloud for cloud-routed access, but that is not local execution. For actual local inference, use llama.cpp with the Unsloth GGUF files.

Which model has the best tool calling for agent workflows?

Claude Sonnet 4.6. It has the most mature implementation with interleaved tool calls during extended thinking, strict JSON mode for validated outputs, 64% lower tool-call latency compared to the previous generation, and the only production-proven computer use capability of the three. GLM 5.2 is strong for coding-specific tool use (file ops, terminal, APIs). MiniMax M3 supports function calling but has the thinnest production track record among the three.

Is MiniMax M3 safe to use with sensitive or proprietary data?

MiniMax is headquartered in Shanghai and operates under Chinese data governance laws including the 2017 National Intelligence Law. If you process sensitive data through the MiniMax API, data governance rules differ from US or EU-based providers. Self-hosting M3 on your own infrastructure using the open weights eliminates the API-based data sovereignty concern, but requires 75 to 150GB of memory and careful license review for commercial deployment.

Which model should I start with if I am building my first agent?

Claude Sonnet 5 is the safest starting point today at $2/$10 per million tokens. It has the strongest instruction following, the most reliable tool use, the best documentation, and the largest ecosystem of integration examples — and it is both cheaper and stronger than Sonnet 4.6, so there is no reason to start a new build on 4.6. Once your agent is working well, test GLM 5.2 or MiniMax M3 on the same tasks to see whether the cost savings justify switching for your specific workload.

Is GLM-5.3-Flash a better deal than MiniMax M3?

On price and licence, yes. GLM-5.3-Flash is $0.15/$0.50 per million tokens with $0.03 cached input, roughly half M3's sub-512K rate and a quarter of what M3 charges above 512K, with no context cliff. It is MIT licensed rather than commercially restricted, it takes image and video input like M3 does, and it carries the same 1M context window in a much smaller 320B/18B-active package. M3's remaining advantages are its sparse-attention behaviour at the far end of the window and a longer track record in agent tooling. For most new cost-sensitive agents, Flash is the one to test first.

How much does it cost to run an AI agent per day?

At 500 tasks a day with 3,000 input and 800 output tokens per task, and no caching: $0.43/day on GLM-5.3-Flash, $0.93 on MiniMax M3, $3.86 on GLM 5.2 or 5.3, $7.00 on Claude Sonnet 5, $10.50 on Sonnet 4.6 and $14.00 on Opus 5.5. With routing (65% to M3, 25% to GLM 5.2, 10% to Sonnet 5), a 2,000-task day lands at about $9.08, roughly $272 a month. The cost section above has the full tables.

Does prompt caching close the gap between Claude and the cheap models?

Partly. A Claude cache hit costs a tenth of base input on Sonnet, so on a task where 80% of the input is repeated context, Sonnet 4.6 drops from $3.00 to $1.92 and Sonnet 5 from $2.00 to $1.28. But output is never cached, and output is where most agent spend goes. Caching narrows the input gap a lot and barely touches the output gap. The cheap models cache cheaply too: $0.26/M cached input on GLM and $0.06/M on M3.

What is the best LLM for an OpenClaw agent in 2026?

There isn't a single one. For customer-facing and high-reliability work, Claude Sonnet 5 at $2/$10 is the default, with Opus 5.5 at $4/$20 as the escalation for the hardest reasoning. For long-horizon autonomous coding, GLM 5.3 at $1.40/$4.40 is the strongest price-to-performance option with open weights. For high-volume cheap decisions, GLM-5.3-Flash at $0.15/$0.50 wins on cost, and MiniMax M3 at $0.30/$1.20 is the pick when you need video input. Most production agents should route between them by task.

Why do different sites quote such different prices for these models?

Most of them quote resellers rather than labs. Z.ai's own rate for GLM 5.1, 5.2 and 5.3 is $1.40/$4.40, while aggregators list the same models at $0.98 or $1.00 depending on which provider is cheapest that week. MiniMax's own rate for M3 is $0.30/$1.20 under 512K input, but many pages print the above-512K tier of $0.60/$2.40 and label it as standard. Budget against the lab's published pricing page and treat any cheaper reseller rate as upside.

One config, every model.

Connect GLM 5.2, Claude Sonnet 4.6, and MiniMax M3 through BetterClaw with BYOK. Test them side by side on your real workload. Free forever, not a trial. Start free →