Most LLM providers now market at least one model for software engineering, but the coding claims, prices, context windows and access terms are spread across dozens of pages and are hard to line up. This page compares the models that eight providers position for coding on their own official pages, using identical fields for each: API identifier, release date, context window, maximum output, price per million tokens, access status and the provider's own coding wording. It covers general frontier models that providers describe for coding, such as Claude Opus 5.5, Gemini 4 Argon and GPT-6.1 Sol, and dedicated coding models such as Codestral, Grok Build, MAI Code and North Mini Code. It is written for engineering leads, platform teams and buyers choosing a model for coding assistants and coding agents. Providers are listed alphabetically; the order is never a ranking, and the page draws no verdicts.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. Model line-ups and prices change often, so confirm against each provider's current documentation.
Key takeaways
- Two kinds of coding model exist. Some providers position a general frontier model for coding (Claude Opus 5.5, Gemini 4 Argon, GPT-6.1 Sol, Grok 4.7, Muse Wonder); others ship a dedicated coding model (Codestral 25.08, Grok Build 0.1, MAI Code 1.1 Flash, North Mini Code 1.0).
- Context windows range from 256K to about 1M tokens. Grok Build 0.1 and North Mini Code 1.0 list 256K, Grok 4.7 lists 500K, and Claude Opus 5.5, Gemini 4 Argon, GPT-6.1 Sol and Muse Wonder list about 1M.
- Prices and rules differ. Listed input prices for the API models run from $1.00 (Grok Build 0.1 under 200K prompt tokens) to $10.00 (GPT-6 Astra) per million tokens; Gemini 4 Argon's price is scheduled to double on 1 December 2026.
- Some coding models are in early access or preview. Grok Build 0.1 is in early access and Muse Wonder is in preview; North Mini Code 1.0 is also listed on OpenRouter as a free variant.
- Few official pages publish coding scores. Of the official model, pricing and changelog pages checked, only Microsoft's MAI Code page published coding benchmark figures; treat any provider score as the provider's own test.
Which models count as coding LLMs?
Almost every current LLM can write code, so this page uses an objective inclusion rule: a model is included when its provider's own model page, model list or changelog describes it for coding or software engineering, and its facts could be confirmed on official pages on 2026-11-15. The provider wording, from the pages checked:
- Anthropic, Claude Opus 5.5: "a model for long-running agentic coding and knowledge work" (release notes); the models overview recommends it as the default for most workloads.
- Cohere, North Mini Code 1.0: "a 30B total / 3B active parameter MoE model trained for agentic coding".
- Google, Gemini 4 Argon: an efficient frontier reasoning model; the model page lists complex coding, agentic, maths and science tasks.
- Meta, Muse Wonder: Meta's most capable model, "for complex software engineering and long-horizon agentic work", in preview. Muse Spark lists software engineering assistants as a use case.
- Microsoft, MAI Code 1.1 Flash: a fast, efficient coding model for GitHub Copilot, built for agentic coding in real repositories.
- Mistral AI, Codestral 25.08: "Our cutting-edge language model for code completion." Mistral Small 4 is described as unifying instruct, reasoning and coding.
- OpenAI, GPT-6.1 Sol and GPT-6 Astra: the changelog records GPT-6.1 Sol as released "for complex coding and professional work at a lower cost than GPT-6 Astra"; the models page calls GPT-6 Astra "our flagship model for complex reasoning and coding".
- xAI, Grok Build 0.1 and Grok 4.7: the release notes describe Grok Build 0.1 as a coding model "trained specifically for agentic coding workflows" and Grok 4.7 as a frontier model "for coding, agentic tasks, and knowledge work".
Providers whose official pages checked on 2026-11-15 did not describe a current model specifically for coding are not listed; see each provider's hub page for its full line-up, for example DeepSeek Models, Alibaba Qwen Models and Zhipu Z.ai GLM Models.
Key facts side by side
As of 2026-11-15. Providers alphabetical. "Not stated" means the provider did not state the fact on the official pages checked on that date.
| Provider | Model | API ID | Released | Context | Max output | Access status |
|---|---|---|---|---|---|---|
| Anthropic | Claude Opus 5.5 | claude-opus-5-5 | 22 September 2026 | 1M | 128K | Generally available |
| Cohere | North Mini Code 1.0 | north-mini-code-1.0 | Not stated by Cohere on the models page | 256K | 64K | Live on the Chat endpoint; Model Vault deployment; OpenRouter free variant |
| Gemini 4 Argon | gemini-4-argon | Not stated on the model page (latest update October 2026) | 1,048,576 (unified input and output) | 65,536 by default; up to 1,048,576 with long decode | Stable | |
| Meta | Muse Wonder | muse-wonder-1.0 | Not stated on the models page | 1,048,576 | Not stated by Meta | Preview; Meta states availability can vary by country |
| Meta | Muse Spark | muse-spark-1.3 (recommended) | Not stated on the models page | 1,048,576 | Not stated by Meta | Available on Meta Model API |
| Microsoft | MAI Code 1.1 Flash | Not stated by Microsoft (used inside GitHub Copilot) | Not stated on the model page | Not stated by Microsoft | Not stated by Microsoft | Available through GitHub Copilot; local runs on supported devices |
| Mistral AI | Codestral 25.08 | Not listed for this version on the overview page | End of July 2025 (v25.08) | Not stated on the overview page | Not stated by Mistral | Listed under current code models |
| Mistral AI | Mistral Small 4 | mistral-small-2603 | 16 March 2026 | 256K | Not stated by Mistral | Generally available; open weights (Apache 2.0) |
| OpenAI | GPT-6.1 Sol | gpt-6.1-sol | 29 September 2026 | 1.05M | 128K | Generally available |
| OpenAI | GPT-6 Astra | gpt-6-astra | 3 September 2026 | 1.05M | 128K | Generally available |
| xAI | Grok Build 0.1 | grok-build-0.1 | 19 May 2026 | 256K | Not stated by xAI | Early access |
| xAI | Grok 4.7 | grok-4.7 | 21 September 2026 | 500K | Not stated by xAI | Available on the xAI API |
Knowledge cutoffs stated on the pages checked: Claude Opus 5.5, June 2026 (reliable knowledge cutoff); GPT-6.1 Sol and GPT-6 Astra, 30 April 2026. The other providers did not state a cutoff for these models on the pages checked. Input types stated: Claude Opus 5.5 text and image; Gemini 4 Argon text, image, video, audio and PDF; Muse Wonder and Muse Spark text, image, video, audio and PDF; Grok 4.7 text and image.
How much do coding LLMs cost?
As of 2026-11-15. USD per million tokens, standard list prices from each provider's official pricing page. Providers alphabetical.
| Provider | Model | Input | Cached input | Output | Pricing rule |
|---|---|---|---|---|---|
| Anthropic | Claude Opus 5.5 | $4.00 | $0.20 | $20.00 | Batch $2 input, $10 output; full 1M window at standard price |
| Cohere | North Mini Code 1.0 | Not stated on the models page | Not stated on the models page | Not stated on the models page | OpenRouter lists a free variant, cohere/north-mini-code:free |
| Gemini 4 Argon | $2.00 to 30 Nov 2026; $4.00 from 1 Dec 2026 | $0.10; $0.20 from 1 Dec 2026 (implicit caching only) | $10.00; $20.00 from 1 Dec 2026 | Batch $1.00 input, $5.00 output to 30 Nov 2026; not on the free tier | |
| Meta | Muse Wonder | $4.00 | $0.20 | $20.00 | Standard mode only |
| Meta | Muse Spark (1.3, 1.2, 1.1) | $1.25 | $0.15 | $4.25 | No long-context premium; discounted Contributor mode on 1.3 and 1.2 |
| Microsoft | MAI Code 1.1 Flash | Not stated as a token price | Not stated | Not stated as a token price | Microsoft states no per-inference charges for local model calls on supported devices |
| Mistral AI | Codestral 25.08 | Not stated on the overview page | Not stated | Not stated on the overview page | See Mistral's pricing page |
| Mistral AI | Mistral Small 4 | See rule | Not stated on the model card | See rule | The model card lists $0.15 and $0.60 per million tokens without labelling which is input and which is output |
| OpenAI | GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | Standard tier price |
| OpenAI | GPT-6 Astra | $10.00 | $1.00 | $50.00 | Standard tier price |
| xAI | Grok Build 0.1 | $1.00; $2.00 at 200K prompt tokens and above | $0.20; $0.40 | $2.00; $4.00 | Long-context rate applies to all tokens once a prompt reaches 200K |
| xAI | Grok 4.7 | $2.00; $4.00 at 200K and above | $0.50; $1.00 | $6.00; $12.00 | Grok 4.7 Fast, the same model, costs 2x (1.5x for long context) and is offered only in Cursor and Grok Build |
Coding agents read large parts of a repository and re-send context on every turn, so cached input prices and long-context thresholds often matter more than the headline input price. For how caching and long-context tiers work, see LLM Pricing Explained and the LLM Pricing Tracker.
What is the context window of each coding model?
Context size decides how much of a codebase, test output and conversation history a model can consider in one request.
- About 1M tokens: Claude Opus 5.5 (1M), Gemini 4 Argon (1,048,576, shared between input and output), GPT-6.1 Sol and GPT-6 Astra (1.05M), Muse Wonder and Muse Spark (1,048,576).
- 500K tokens: Grok 4.7.
- 256K tokens: Grok Build 0.1, Mistral Small 4, North Mini Code 1.0.
- Not stated on the pages checked: Codestral 25.08, MAI Code 1.1 Flash.
Maximum output limits how large a single generated change or file can be: Claude Opus 5.5, GPT-6.1 Sol and GPT-6 Astra list 128K, North Mini Code 1.0 lists 64K, and Gemini 4 Argon lists 65,536 by default with long decode up to the full window. See LLM Context Windows Compared Across Providers.
Coding features stated by providers
As of 2026-11-15. Features and tools are listed only where the provider's page states them for the model. Providers alphabetical.
| Provider | Model | Coding-related features stated |
|---|---|---|
| Anthropic | Claude Opus 5.5 | Long-running agentic coding; tool use; Batch API at 50% of standard rates |
| Cohere | North Mini Code 1.0 | Agentic coding; mixture-of-experts (30B total, 3B active); Model Vault production deployment |
| Gemini 4 Argon | Thinking (low, medium, high), function calling, structured outputs, code execution, computer use, file search, URL context | |
| Meta | Muse Wonder | Makes and verifies changes across many files in real repositories; Muse Wonder Full variant in Muse Code |
| Microsoft | MAI Code 1.1 Flash | Agentic coding, repository questions and refactoring; native VS Code and Copilot CLI integration; languages listed: Python, TypeScript, Java, JavaScript, HTML5, .NET, CSS3, C++ |
| Mistral AI | Codestral 25.08 | Code completion |
| OpenAI | GPT-6.1 Sol | Complex coding and professional work |
| xAI | Grok Build 0.1; Grok 4.7 | Agentic coding workflows (Grok Build); reasoning effort low, medium, high (default) and xhigh (Grok 4.7) |
Benchmarks
Coding benchmarks such as SWE-bench Verified and Terminal-Bench measure whether a model can resolve real repository issues or complete terminal tasks, but scores depend on the agent harness, tool access, number of attempts and benchmark version. Provider-reported figures and independent leaderboards are kept separate below.
| Provider | Model | Provider-reported coding score (source, as of) | Independent leaderboard figure |
|---|---|---|---|
| Anthropic | Claude Opus 5.5 | Not stated on the pages checked | Not listed on this page |
| Cohere | North Mini Code 1.0 | Not stated on the pages checked | Not listed on this page |
| Gemini 4 Argon | Not stated on the model page | Not listed on this page | |
| Meta | Muse Wonder | Not stated on the pages checked | Not listed on this page |
| Microsoft | MAI Code 1.1 Flash | SWE-Bench Verified 72.6% pass rate; Terminal Bench 2.1 62.9% pass rate (Microsoft MAI Code model page, checked 2026-11-15; no harness settings stated) | Not listed on this page |
| Mistral AI | Codestral 25.08 | Not stated on the overview page | Not listed on this page |
| OpenAI | GPT-6.1 Sol | Not stated on the pages checked | Not listed on this page |
| xAI | Grok Build 0.1 | Not stated on the release notes | Not listed on this page |
The SWE-bench project publishes official Verified, Lite, Multilingual, Multimodal and Bash Only leaderboards at swebench.com; check the date of any figure you take from it, and compare models only on the same leaderboard and setup. See How to Read LLM Benchmarks.
Deprecations and retirements of coding models
Dedicated coding models have short lifecycles. From the official pages checked on 2026-11-15:
- Mistral AI: every Devstral version is deprecated. Devstral 2 (devstral-2512) was deprecated on 22 May 2026 and retired on 31 July 2026, with Mistral Medium 3.5 as the replacement; Devstral Small 1.1 (devstral-small-2507) retired on 31 May 2026 with Mistral Small 4 as the replacement. Codestral 25.01 (codestral-2501) retired on 30 November 2025.
- OpenAI: the changelog records gpt-5.2-codex (14 January 2026) and gpt-5.3-codex (24 February 2026) releases to the Responses API; current coding positioning sits with GPT-6.1 Sol and GPT-6 Astra, and no Codex model is priced on the pricing page checked.
- xAI: Grok Build 0.1 is in early access, so its identifier and pricing may change before general availability.
- Microsoft: MAI Code 1.1 Flash is the current version on the model page, with MAI Code 1 Flash shown as the earlier release.
Pin model identifiers in coding tools and CI, and track each provider's deprecations page. See the LLM Deprecation and Retirement Tracker.
Data handling and security for coding models
Coding models see source code, configuration files and often secrets, and coding agents can run commands. The model choice is only part of the control set.
- Source code exposure: confirm each provider's retention and training terms before sending proprietary code; local options, such as MAI Code 1.1 Flash on supported devices or self-hosted open weights like Mistral Small 4, keep inference on your own hardware.
- Secrets in context: repository context sent to a model can include keys and tokens; scan repositories and block secrets from prompts.
- Agent permissions: coding agents that can execute code or use computer-use tools need scoped credentials, sandboxing and human approval for sensitive actions.
- Keys and spend: agent loops can consume tokens fast; set spend limits and alerts on every key.
See the AI IDE and Coding Agent Security Guide, AI Coding Tools Secrets Sprawl, the Vibe Coding Security Guide and How to set spend limits on LLM API keys.
How to choose a coding model: factual criteria
- Task type: code completion in an editor (Codestral is described for completion) differs from multi-file agentic changes (the agentic coding models above).
- Repository size: match context window and long-context pricing to the amount of code an agent reads per task.
- Token pattern: agents re-send context, so compare cached input prices and 200K thresholds (Grok models) as well as base prices.
- Scheduled price changes: Gemini 4 Argon prices double on 1 December 2026.
- Deployment: local or self-hosted options exist for MAI Code 1.1 Flash (supported devices) and Mistral Small 4 (open weights).
- Maturity: early access (Grok Build 0.1) and preview (Muse Wonder) status can mean changing behaviour and identifiers.
- Your own evaluation: run the same set of real tasks from your codebase on each candidate, with the same agent harness, before choosing.
For tool-level selection, see the AI Coding Assistant Buyer's Guide.
Practitioner checklist
- Record every coding model in use, with API ID, provider and price date, in the AI inventory.
- Evaluate candidates on your own repository tasks with one fixed harness before switching.
- Model cost per task, including cached context and long-context thresholds, not just per token.
- Pin model identifiers in IDE settings, agents and CI, and watch deprecation pages.
- Confirm retention and training terms before sending proprietary source code.
- Block secrets from prompts and scan repositories used as agent context.
- Give coding agents least-privilege credentials, sandboxed execution and approval steps.
- Set spend limits and alerts on API keys used by agents.
Standards and references
- Anthropic: Models overview (accessed 2026-11-15)
- Anthropic: Pricing (accessed 2026-11-15)
- Anthropic: Claude Platform release notes (accessed 2026-11-15)
- Cohere: Models overview (accessed 2026-11-15)
- Google: Gemini 4 Argon model page (accessed 2026-11-15)
- Google: Gemini API pricing (accessed 2026-11-15)
- Meta: Muse models (accessed 2026-11-15)
- Meta: Model API pricing and rate limits (accessed 2026-11-15)
- Microsoft AI: MAI Code model page (accessed 2026-11-15)
- Mistral AI: Models overview (accessed 2026-11-15)
- Mistral AI: Mistral Small 4 model card (accessed 2026-11-15)
- OpenAI: Models (accessed 2026-11-15)
- OpenAI: Pricing (accessed 2026-11-15)
- OpenAI: API changelog (accessed 2026-11-15)
- SWE-bench: Official leaderboards (accessed 2026-11-15)
- xAI: Pricing (accessed 2026-11-15)
- xAI: Release notes (accessed 2026-11-15)
Related NHI Mgmt Group resources: AI Coding Assistant Buyer's Guide, AI IDE and Coding Agent Security Guide, Small and Low-Cost LLMs Compared, Frontier AI Models Compared, How to Read LLM Benchmarks.