Codestral 2 (Mistral)
B Tier · 7.5/10
Mistral's dedicated code model -- Codestral 2 (launched 2026-04-08) relicensed under Apache 2.0, removing the commercial-use restrictions of the original. 22B dense, strong FIM (fill-in-middle), available via Mistral API + Hugging Face
Score Breakdown
The Good and the Bad
What we like
- +Relicensing to Apache 2.0 is the real news -- the original Codestral required a Mistral Non-Production license for any commercial use, which blocked adoption in-product. Codestral 2 is immediately usable in commercial IDEs, coding assistants, and CI tooling
- +FIM (fill-in-middle) performance is class-leading for open models -- purpose-built for IDE autocomplete in a way that general-purpose models (Llama, DeepSeek V3) aren't. Competitive with GitHub Copilot's underlying model for inline completions
- +22B dense (not MoE) means predictable VRAM requirements and throughput -- easier to deploy than DeepSeek's 671B MoE or Qwen's sparse 35B-A3B for teams that want certainty
- +Available via Mistral's EU-hosted API for customers who need GDPR-native inference -- rare combination of 'open weights + EU vendor' in the code-model category
What could be better
- −22B parameters put it behind frontier closed models (Claude Opus 4.7, GPT-5.4, Gemini 3.1 Pro) on complex multi-file reasoning and agentic coding. This is a fast, cheap inline-completion model, not a frontier coding agent
- −No multimodal or tool-use baked in -- if your workflow needs screenshot-to-code or terminal tool execution, Claude Code, Cursor Composer 2, or Devin cover that ground better
- −Benchmark transparency could be stronger -- Mistral publishes MBPP / HumanEval numbers but third-party SWE-bench or LiveCodeBench verification is thinner than for DeepSeek, Qwen Coder, or the frontier models
- −SWE-bench Verified performance trails the top open-weight coding specialists (Qwen Coder 3.5, DeepSeek V3 Coder variants) by several points in independent testing
Pricing
Open weights (Apache 2.0)
- ✓22B dense model on Hugging Face
- ✓Commercial use allowed (new in Codestral 2; original Codestral required Mistral Non-Production license)
- ✓Self-host on your own infrastructure
- ✓Fine-tune without license fees
Mistral La Plateforme (hosted API)
- ✓Pay-as-you-go API access
- ✓FIM (fill-in-middle) endpoint for IDE autocomplete
- ✓Chat + completion endpoints
- ✓Consistent with Mistral Small/Medium tier pricing
Self-hosted (Hardware)
- ✓Min: 48 GB VRAM (1x RTX 6000 Ada or 2x RTX 3090 with tensor parallelism)
- ✓Mid: 1x H100 80GB for production throughput
- ✓Max: 2x H100 for batched serving + low latency
- ✓Quantized (GGUF Q4_K_M) runs on a 24GB card for experimentation
System Requirements
Hardware needed to self-host. Min = smallest viable setup (usually heavy quantization). Max = full-precision / production-grade.
| Model variant | Min | Max |
|---|---|---|
| Codestral 2 22B dense (Apache 2.0)Apache 2.0 commercial use OK. Original Codestral (2024) still under Mistral Non-Production License -- verify you are on Codestral 2. | 48 GB VRAM -- 1x RTX 6000 Ada or 2x RTX 3090 tensor parallel (or quantized GGUF Q4_K_M on a 24GB card) | 1x H100 80GB for production FP16 throughput; 2x H100 for batched serving |
Known Issues
- Codestral 2 is Apache 2.0, but the ORIGINAL Codestral (2024) is still under Mistral Non-Production License -- if you pulled older weights before 2026-04-08, verify you're on Codestral 2 before shipping commercial useSource: Mistral release notes · 2026-04
- EU-hosted API infrastructure can have higher latency than US-based DeepSeek or GitHub Copilot backends for North American developersSource: Developer reports on Mistral Discord · 2026-04
Best for
Developers and teams who want a legally-clean open-weights code model they can self-host OR hit via API, particularly those with EU data-residency requirements. Ideal for building in-house IDE extensions, code-review bots, or CI/CD AI integrations where the Apache 2.0 license removes procurement friction.
Not for
Developers who want frontier-quality agentic coding -- Cursor Composer 2, Claude Code, or Devin will outperform on complex multi-file tasks. Also not ideal if you only need hosted inference and don't care about self-hosting -- DeepSeek V3.2 and Qwen3.6-Plus offer stronger benchmarks at competitive pricing.
Our Verdict
Codestral 2's Apache 2.0 relicensing is the biggest licensing unlock in open-source coding models since Meta released Llama 2 commercially. The model itself is solid-not-frontier (22B dense, fast, predictable), but the license change is what matters -- teams that couldn't touch the original Codestral because of commercial restrictions can now ship it in products. For IDE-style inline autocomplete on owned infrastructure, or for EU-data-residency use cases, this is now a first-tier option. For agentic or frontier coding work, keep using Claude Opus 4.7 via Claude Code or Composer 2 in Cursor.
Sources
- Mistral news (accessed 2026-04-18)
- ReleaseBot: Mistral updates (accessed 2026-04-18)
- fazm.ai: April 2026 open model releases (accessed 2026-04-18)
Explore more Codestral 2 (Mistral) rankings
Deeper leaderboards, benchmarks, task-specific tier lists, and status/pricing pages for Codestral 2 (Mistral).
The Tier List Tuesday
Weekly newsletter: tier movers, new entrants, and the VS of the week. Built from our daily AI-tool sweeps. No spam, unsubscribe anytime.
Alternatives to Codestral 2 (Mistral)
GitHub Copilot
**Oct 15: @GitHub in Slack and Microsoft Teams now works across multiple repositories in a single conversation (issues, pull requests, bug investigations), can be mentioned anywhere in a message, and can use shared Slack messages and threads as context.** **Gemini 4 Argon joined Copilot on 2026-10-13, the day Google made it GA, for Pro+, Max, Business and Enterprise at provider list price; on 10/15 GitHub scheduled Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra for removal from every Copilot surface on 2026-11-16 (Opus 4.8 stays for individual annual subscribers), and HydraFusion gained a Max option and a pool of Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna (10/14).** **Claude Haiku 5.5 joined Copilot on 2026-10-07, the day Anthropic launched it, for Pro, Pro+, Max, Business and Enterprise at provider list price; the same day local sandboxing went GA in Copilot CLI, the Copilot app and VS Code Agent Host at no extra cost, the CLI learned to discover local Ollama models, and GitHub announced a purpose-built secret-detection model whose new push-protection and /security-review checks will draw AI Credits.** **GPT-6.1 Sol landed in Copilot on OpenAI's launch day (2026-09-29, Pro+/Max/Business/Enterprise), the fifth day-one frontier add in eight days, and on 2026-10-02 the four-model retirement slate executed on schedule: Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code and Claude Opus 4.7 are gone from every Copilot surface.** Same week: HydraFusion's multi-model orchestration reached VS Code 1.140 and the Copilot app (9/30), **computer use** let Copilot CLI and the Copilot app drive desktop applications on macOS and Windows in public preview (10/01), dynamic workflows defined in code arrived for CLI, app and SDK (10/01), the September VS Code roundup added automations, agent merge and Dev Container sessions (10/01), and Copilot code review gained REST/GraphQL request support with Balanced as the default effort level (10/02). Still ahead: the six-model slate on 10/19 and the 10/22 default policy for new features for Business and Enterprise. The 9/22 and 9/28 adds (Opus 5.5, GPT-6 Sol/Luna, Sonnet 5.5) all stand, metered at provider list pricing.
Cursor
**Oct 16: cloud agents can now build interactive canvases and share them by link, the Environments API entered beta for creating and managing agent environments in code, and cloud agent environments hold up to 300 repositories with large environments loading 5x faster.** **Cursor's own plans closed to new customers on 2026-10-12: Cursor is now sold through The X Subscription (Free, Plus $8, Premium $30, Super $100, Ultra $200) with one usage pool shared with Grok, at least twice the coding usage on SpaceXAI's models -- but Claude, GPT and Gemini are no longer in the allowance, billed at API list price on Premium and up and unavailable on Plus. On 2026-10-06 the Cursor iOS app gained remote control of local agents.** **Cursor launched two bots for the last mile of shipping code on 2026-09-23 (Teams and Enterprise): Rollouts, which attaches a monitor to every pull request and watches the change as it deploys per environment -- verified healthy, regression detected, or inconclusive -- and Security Review, which posts one comment per PR reporting exploitable bugs with severity, attack path and a proposed fix.** Rollouts is the Cursor rebuild of Firetiger's Change Monitors on the Bot Development Kit; it can open a revert PR or hand a regression to a cloud agent but does not merge or roll back on its own. Models: Anthropic's 9/28 Sonnet 5.5 post quotes SpaceXAI's Sualeh Asif on CursorBench 4.0 -- Sonnet 5.5 at 55.5% 'second only to Opus 5.5' (57.8%); Grok 4.7 and its Cursor-only Fast variant arrived 9/21. **Cursor Projects (9/10, beta)** and the 11/12 OpenAI shutoff watch continue.
Devin Desktop (formerly Windsurf)
Windsurf is now **Devin Desktop** -- Cognition retired the Windsurf brand via OTA update on June 2, 2026. Same editor, plans, pricing, settings, and extensions; the bundled agent is now 'Devin Local' and Devin Cloud agent access starts on the $20 Pro plan. Agent Command Center, Spaces, and Devin Review all carry over
Tabnine
**Tabnine was acquired by Tricentis on 2026-07-30 and is being folded into the Tricentis Agentic Quality Engineering Platform as its Enterprise Context Engine; tabnine.com now redirects to tricentis.com and the public price list is gone, though the VS Code extension (v3.335.0, about 9.6M installs) and the docs are still live.** Privacy-first AI code completion with on-premise deployment -- now a testing-company asset rather than a standalone Copilot rival.
Claude Code
**2.1.302 (2026-10-16) added CLAUDE_CODE_AUTO_MODE_MODEL to choose the model behind the client-side auto-mode classifier, an Idle compaction on/off row in /config, and a notice in /design-sync and /design-login that standalone Claude Design (claude.ai/design) closes by 2026-12-14; it also fixed file writes on exFAT/FAT32 drives on Windows and stopped prompt suggestions spending credits.** **2.1.299 (2026-10-14) added idle compaction -- a 200K+ token conversation left idle about 54 minutes is compacted before its one-hour prompt cache expires -- plus /effort and /fast defaults in agent view and /usage warnings for max effort and large tool results; 2.1.300 (10/14) brought /advisor to Claude Platform on AWS; 2.1.301 (10/15) lets unattended retry mode give up on far-off usage-limit resets and makes the fable alias fall back to the previous Fable model.** **Claude Haiku 5.5 became the default Haiku in 2.1.293 (2026-10-07) at $0.10 / $0.50 per 1M below 100K tokens; 2.1.295 (10/08) added onFailure block for hooks, OSC 7501 terminal status, per-upstream model lists on the Claude apps gateway and a 16,384-character limit on tool-search MCP descriptions, and 2.1.296 (10/09) added an allow_large Read option, per-subagent autoCompactWindow and a code policy key for Claude Desktop's Code tab.** **Three releases in two days -- 2.1.290 (2026-10-05), 2.1.291 and 2.1.292 (both 10/06): claude attach and claude logs by session name, a /claude-api managed-agents-onboard command, an effort parameter on the Agent tool, --marketplace on claude plugin install, a configurable 529 retry delay, WebFetch continuation past 100,000 characters, local MCP servers defaulting to protocol 2026-07-28, and a long run of security fixes for symlink and UNC-path reads; no model or pricing changes.** Anthropic's terminal-based coding agent that reads your whole repo and makes real changes -- not just suggestions. **Claude Opus 5.5 became the default Opus model in 2.1.280 (2026-09-22) and Claude Sonnet 5.5 the default Sonnet in 2.1.284 (2026-09-28)**: Opus 5.5 at $4/$20 with $0.20 cache reads and fast mode at $8/$40 for up to 2.5x speed, Sonnet 5.5 at an unchanged $2/$10 with far fewer tokens per task. Both are 1M context. Fable 5.1 (2026-09-01) remains the top model and still defaults to High effort in Claude Code specifically. **Four more releases 9/29-10/02 (2.1.285-2.1.288) added Claude Mods (plugins that modify deeper behaviour) with a built-in 'You should know' side agent, `claude --desktop` to hand a session to the desktop app, an `allowedProviders` managed setting, a WebFetch kill switch, `--max-findings` on /code-review, and continuation from partial responses when the API times out mid-turn.** Also since 9/22: `/doctor prompt-audit`, `deniedModels`, dollar amounts on spend limits and `attribution: false`.
Lovable
**Lovable launched Growth tools on 2026-10-14: set a goal such as your first 100 signups and it builds a plan from your app and traffic, prepares SEO, listings and creative, and launches Google, YouTube, Facebook and Instagram ad campaigns you approve, with no extra Lovable fee. Same week: Sign in with Lovable for any website, 60-day restore for deleted projects, Gemini 4 Argon (10/13) and GPT-6.1 Sol (10/15) for app AI features, and Enterprise audit-log export to SIEMs.** **Lovable shipped credit usage limits and alerts for workspaces (10/07), invite-only publishing on Free and Pro (10/08), nine new app connectors plus a Loom chat connector, Drizzle migration files for Cloud database changes, chat-to-chat task handoff and Enterprise workspace publishing domains (10/09).** Describe the app you want in plain English and watch it build itself. **New this week: apps built on Lovable can generate music with Google's Lyria 3 Clip Preview (about 30-second clips) and Lyria 3 Pro Preview (tracks up to 184 seconds) from a prompt or an image (2026-10-05, credits charged per finished track), beehiiv and Ramp connectors arrived, Lovable can restore its own work after a Git push overwrites it, and GitHub secret scanning now detects leaked Lovable API keys (10/05). Earlier: live voice conversations through GPT Live (2026-10-01, TanStack Start apps, uses credits), Enterprise workspaces can pin AI processing to EU providers (10/01, costs more credits), Business and Enterprise API keys can be IP-allowlisted (10/01), Lovable can rotate a Cloud database password from chat (9/30), and the Lovable MCP server works inside Muse by Meta (9/29).** **Chat with Lovable (2026-09-21)** adds a free daily chat allowance for thinking, asking and planning without changing code, plus voice conversations and a Telegram surface; apps built on Lovable can now call **Claude Opus 5.5 (9/23)** and **GPT-6 Sol and Luna (9/25)** for their own AI features, and the **Lovable API (9/17)** automates workspaces. $500M ARR (June 2026) and ~1M new projects a week
Devin
**Oct 16 release notes: /devin review and /devin prompts now work on GitLab merge requests, Devin Review shows Bitbucket Cloud merge readiness, Devin gained built-in git tools for issues, labels, assignees and branch protection plus PR search across GitHub, GitLab, Azure DevOps, Gerrit and Perforce Swarm, and queued messages stay editable while Devin is idle.** **Devin now has memory: the 2026-10-09 release notes add personal memory across sessions with a nightly 'dreaming' pass that refines what it keeps (admins can force it on or off), cut the session resume window from 30 to 14 days, load a repository's AGENTS.md on setup and open Devin Ultra to zero-data-retention enterprises; 10/07 added a notification inbox. Anthropic's Haiku 5.5 launch lists it as a Devin Fusion sidekick option.** **10/02 and 10/05 release notes: a Terminal tab in the session workspace, keyboard shortcuts and a command palette for everyone, Devin Review effort levels set per trigger with an enterprise default, single-code-scan fetch via API and MCP, Azure DevOps review summaries posted as closed threads, Perforce snapshot builds given three hours, and Microsoft Teams access settings now enforced.** **9/30 release notes: Devin Review now routes findings on a Devin-opened PR back to the session to fix first (only what it cannot fix gets posted), automations gain preflight scripts, and dropped-connection messages are no longer lost.** **Cognition crossed a one billion dollar run rate on 2026-09-25** -- '$1B in annualized revenue run rate', less than two years after Devin's GA, naming GE Aerospace, Rivian, Rohlik and Exa -- and opened a Sao Paulo hub on 9/22 with Itau (over 75% of its technology teams on Devin, 300,000+ repositories documented, .NET-to-Java migrations 6x faster), Nubank (a multi-million-line monolith migration 'from years to weeks' at over 20x lower cost), Santander, Natura and EBANX as customers. **Devin Fusion (9/11)** -- the lead/sidekick two-model harness measured at 23-46% lower cost per task -- and the 9/15 AWS multi-year agreement stand.
Replit
Cloud IDE with an AI agent that builds full apps from prompts. **Free Mode (2026-08-18)**, powered by OpenAI's GPT-5.6 Luna, lets Core and Pro subscribers chat and run everyday tasks without spending credits -- up to 30 hours of chat a month on Core -- and **Intelligent Model Routing (2026-08-26)** now picks the model per task at what Replit measured as 65% lower cost than the old Max Mode. Replit acquired the charting startup Atta on 2026-09-25. Agent 4 (May 2026) added parallel task execution
Codex (OpenAI)
**GPT-6.1 Sol Ultrafast is now on the API rate card -- $12 input / $0.60 cached / $15 cache writes / $60 output per 1M up to 272K tokens, six times Standard, closing the DevDay 'coming days' promise -- and on 2026-10-09 the Codex desktop app gained Composer predictions (beta), which suggests your next message after each response for personal Pro users on GPT-6 Astra and GPT-6.1 Sol.** **OpenAI and Atlassian expanded their partnership on 2026-10-06: GPT-6 family models (GPT-6 Astra and the GPT-5.6 series are named) will power agents across Atlassian's platform and Rovo, more than 3,000 Atlassian developers already use Codex, and the Atlassian and Teamwork Graph CLI plugins connect Codex and ChatGPT to Jira, Confluence, Looms and Bitbucket pull requests.** **Two October notes: EU Codex output will carry OpenAI's invisible textGrain text watermark within weeks (API customers can opt in worldwide from 10/05), and the new Enterprise token rate card prices Codex at API Standard rates, names GPT-5.3-Codex as the code-review model and GPT-5.6 Luna as the auto-review model, and puts Astra Ultrafast at 6x.** GPT-6.1 Sol Ultrafast is still not on the API rate card as of 10/05. **DevDay 2026 (2026-09-29) was mostly a Codex release: GPT-6.1 Sol in Codex and the API at $2/$10 per 1M with cached input cut to $0.10, Codex Cloud (reusable cloud environments, tasks that keep running while your laptop sleeps, on Plus and up), a refreshed CLI with voice steering and an /agents view, Code Review in the desktop app with automatic cloud first passes, Codex Security Cloud with Daybreak Blue models and no separate Daybreak application, and the Agents API gaining computer use** -- plus GPT-6 Astra Ultrafast (300 tokens per second in Codex) on Pro 500 and Enterprise, with GPT-6.1 Sol Ultrafast 'coming soon'. GPT-6.1 Sol nearly matches GPT-6 Astra on DeepSWE at one-fifth of the token price and also landed in GitHub Copilot the same day. Still in place: GPT-6 Sol and Luna (9/22), the Agents API public beta (9/10) and the 8/31 retirement of GPT-5.4 from Codex. On 10/01 OpenAI scheduled GPT-5.3-Codex for API removal on 2027-04-01.
Google Antigravity
**Gemini 4 Argon reached Antigravity on 2026-10-13 for Google AI Ultra subscribers and enterprises, and Google is offering 10,000 scientists Antigravity's highest usage tier free for six months.** Google's agent-first AI IDE -- deploys up to 5 autonomous coding agents in parallel on a VS Code fork. Antigravity 2.0 (I/O 2026) is the runtime substrate for Gemini Spark, and the Antigravity CLI is now the official successor to Gemini CLI, which stopped serving consumer tiers on 2026-06-18
Roblox Assistant
Roblox Studio's agentic AI that plans, builds, and playtests games. Planning Mode (2026-04-16) + Mesh Generation + Procedural Models brings 3D-native creation to 70M+ daily creators