TL;DR: Mistral Large 4 (public preview October 6, 2026) is a 1.05T-parameter MoE with 52B active parameters, a 1M-token context window, and image input, priced at $0.68/$2.09 per million tokens for two weeks, then $1.36/$4.18. As a coding backend it scores 28.3% vendor-reported on Terminal-Bench 4 — below Claude Haiku 5.5’s 39.2% at a seventh of the input price. The reasons to pick it are the context window, EU hosting, and the open weights promised for late October.

	Mistral Large 4	Claude Haiku 5.5	Claude Sonnet 5.5
Input / output per MTok	$0.68 / $2.09 launch ($1.36 / $4.18 after ~Oct 20)	$0.10 / $0.50 (≤100K prompt), $0.50/$2.50 above	$2 / $10
Terminal-Bench 4	28.3% vendor, 22.7% independent	39.2% (vendor)	70.6% (vendor)
Context window	1M tokens (docs; 524K listed for the preview endpoint by Artificial Analysis)	1M (price doubles-plus past 100K prompt)	200K
The catch	Mid-pack coding scores at a mid-tier price; preview endpoint, license TBD	Quality ceiling on long multi-file work	Price, and a 200K context ceiling

Honest take: If you’re choosing a coding backend on results per dollar, Haiku 5.5 beats Mistral Large 4 today and it isn’t close. Wire Large 4 into Cline if you specifically need EU-hosted inference, million-token context without a pricing cliff, or image input in an agent loop — otherwise bookmark it and revisit when the open weights land at the end of October.

What exactly did Mistral ship on October 6?

Mistral Large 4 — internal codename “Le Chonk” — entered public preview on October 6, 2026 as Mistral’s new flagship: a granular Mixture-of-Experts model with 1.05 trillion total parameters, 52B active per token, and a 1.6B-parameter vision encoder. Those numbers come from Mistral’s own model card source (the platform-docs-public repository that generates docs.mistral.ai), which also settles a discrepancy floating around launch coverage: several trackers reported 49B active parameters; Mistral’s card says 52B.

The spec sheet, as Mistral documents it:

Spec	Mistral Large 4
Architecture	Granular MoE, 1.05T total / 52B active
Context window	1M tokens
Input	Text + image (1.6B vision encoder)
Output	Text + reasoning (adjustable reasoning_effort parameter)
API model IDs	mistral-large-4, mistral-large-4-0
Features	Function calling, structured outputs, prefix completion, batch API
Status	Public preview, version 26.10
Weights	”Coming soon” — HuggingFace placeholder mistralai/Mistral-Large-4-1T-A52B, no license named yet

One context-window caveat worth knowing before you plan a giant-repo workflow around it: Artificial Analysis lists the preview endpoint at 524K tokens, not the 1M Mistral’s docs claim. Preview endpoints get capped below the model’s architectural limit often enough that this is plausible rather than alarming — but if your use case is “feed it 800K tokens of monorepo,” verify against your own account’s limits before you commit.

Function calling is listed as a first-class capability, which matters more than any benchmark here: without reliable tool calls, a model can’t drive Cline’s edit-and-execute loop at all. The reasoning_effort parameter works the same way as on GLM-5.3 and Claude’s effort dials — one more knob that trades latency for depth without switching models.

How much does Mistral Large 4 cost — and when does the launch discount end?

$0.68 per million input tokens and $2.09 per million output right now, doubling to $1.36/$4.18 when the launch window closes. Mistral’s changelog states the launch pricing is 50% off for two weeks from the October 6 release — so budget for standard rates from roughly October 20. Cached input is $0.07/M during launch, $0.14/M after.

Here is what that does to a realistic month of agentic coding — 40 agent requests a workday averaging 20K input and 1.5K output tokens, or 17.6M input + 1.32M output across 22 workdays (the same workload we costed in our Haiku 5.5 analysis, with every prompt under 100K tokens):

Monthly cost, same workload	Large 4 (launch)	Large 4 (standard)	Haiku 5.5	Sonnet 5.5
Input 17.6M tokens	$11.97	$23.94	$1.76	$35.20
Output 1.32M tokens	$2.76	$5.52	$0.66	$13.20
Total	$14.73	$29.46	$2.42	$48.40

All Anthropic numbers re-verified against claude.com/pricing on October 11, 2026. Two readings of that table:

• Against Sonnet 5.5, Large 4 at standard pricing costs about 60% as much. If its coding quality were within shouting distance of Sonnet, that would be a real deal. It isn’t (next section).
• Against Haiku 5.5, Large 4 costs 6x–12x more on this workload. The one scenario where the gap collapses: prompts consistently above 100K tokens, where Haiku’s pricing jumps to $0.50/$2.50 and launch-priced Large 4 ($0.68/$2.09) is nearly at parity — with no cliff anywhere in its 1M window. If your agent sessions routinely carry 150K+ tokens of context and you refuse to compact, Large 4’s flat pricing is genuinely the simpler bill.

Is Mistral Large 4 actually good at coding?

Mid-pack — and for once the vendor’s own chart says so. Mistral’s launch materials report 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4, with a footnote that the coding scores were evaluated privately by Artificial Analysis before publication. On the same DeepSWE table that Reflection AI publishes, Kimi K3 scores 68.0 and DeepSeek V4.1 Flash 74.2 — both open-weight rivals, both cheaper, both ahead.

The independent data that exists two weeks in points the same direction:

Measurement	Score	Who ran it
DeepSWE v1.1	61.7%	Vendor (via Artificial Analysis private eval)
Terminal-Bench 4	28.3%	Vendor (same footnote)
Terminal-Bench 4	22.73%	Vals.ai, independent
Intelligence Index	38	Artificial Analysis
Output speed	116 tok/s (18.7s to first answer token)	Artificial Analysis, Mistral API

For calibration: Claude Haiku 5.5 posts 39.2% on Terminal-Bench 4.0 (vendor-run) at $0.10/M input, and Sonnet 5.5 posts 70.6% at $2/M. Large 4’s 22.7–28.3% lands below the cheap model, at nearly seven times the cheap model’s input price. No independent SWE-bench Verified score had been published as of October 11 — if one lands meaningfully above the DeepSWE showing, this calculus changes, but you’d be buying on hope today.

Artificial Analysis also flags a trait that costs you real money in an agent loop: verbosity. Running its index took about 200 million output tokens against an 81M median for comparable models. A model that says 2.5x more per task inflates every output-side invoice line, and output is where Large 4’s pricing is least competitive.

Where the model is genuinely interesting is everything around the coding scores: image input in an agentic loop (screenshot-to-fix workflows without switching models), a 1M window with flat pricing, and speed — 116 tok/s is brisk for a frontier-class MoE. It is a capable generalist that codes, not a coding specialist. Mistral’s coding specialist lineup is Codestral 2 and the Devstral 2 open-weight line.

How do you wire Mistral Large 4 into Cline?

Cline has a native Mistral provider — pick Mistral in the provider dropdown, paste an API key from console.mistral.ai, and type the model ID by hand. We checked Cline’s model catalog source on October 11, 2026: the built-in list still tops out at mistral-large-2512 (that’s Large 3), so Large 4 won’t appear as a preset yet. Enter mistral-large-4 as a custom model ID and it routes fine, because Cline’s Mistral provider passes any model string through to the API.

Sanity-check your key and the model ID from the terminal first:

curl -s https://api.mistral.ai/v1/chat/completions \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-large-4",
    "messages": [{"role": "user", "content": "Reply with exactly: ok"}]
  }'

Expected response (trimmed):

{
  "model": "mistral-large-4",
  "choices": [{"message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}],
  "usage": {"prompt_tokens": 10, "completion_tokens": 1}
}

If that returns a 401, your key is wrong; if it returns a model-not-found error, your account may not have preview access enabled — Large 4 is a public preview, and preview models occasionally lag on older account tiers. The gate that matters is function calling, since Cline’s whole file-edit loop runs on it: Mistral documents it as a supported capability on Large 4, but a brand-new preview endpoint is exactly where tool-call formatting bugs show up first, so run one scoped Cline task and watch the tool calls succeed before trusting it with a long session.

One Cline-specific note: Mistral’s API shares a limitation with most OpenAI-compatible endpoints — tool-result messages must be plain strings, so images returned by tools (browser screenshots, for instance) need special handling. Cline ships middleware that rewrites these before they hit Mistral’s converter, so screenshot-driven debugging works; if you’re using a different harness, test that path before relying on it.

Can you use Mistral Large 4 in Cursor?

Yes, but it’s the janky route. Cursor has no native Mistral key slot — the Settings → Models → API Keys panel covers OpenAI, Anthropic, and Google. Reaching Mistral means the Override OpenAI Base URL option pointed at https://api.mistral.ai/v1 with your Mistral key pasted into the OpenAI key field (the label lies; the key goes wherever the override points), then adding mistral-large-4 as a custom model name. Cursor staff have said on the official forum that this override route is supported and staying.

The problem we hit — and that fills the forum threads — is that “OpenAI-compatible” undersells what Cursor actually sends. Cursor’s OpenAI-bound traffic uses the Responses API format rather than plain Chat Completions, and the base-URL override is global, so it hijacks every OpenAI-routed model at once, not just the one you added. In practice requests can fail against a perfectly valid key and endpoint. The fix that works: don’t point Cursor at Mistral directly. Either route through a gateway that explicitly documents Cursor compatibility (OpenRouter carries Large 4 as mistralai/mistral-large-4-0) or — simpler — keep Cursor on its native providers and run Mistral experiments in Cline, where the provider integration is first-class. We covered why Cursor BYOK plumbing breaks this way in the OpenAI-Cursor model access piece.

Should you wait for the open weights instead?

If self-hosting is your angle, yes — wait, but check the license before celebrating. Mistral’s model card marks the weights “coming soon” with a HuggingFace placeholder at mistralai/Mistral-Large-4-1T-A52B; launch coverage puts the release between October 27 and 31, and VentureBeat reports it will ship under a custom Mistral license whose terms haven’t been published. As of October 11 there is no repo live and no license text. “Open-weight” under a restrictive custom license is a very different asset than Apache 2.0 — Kimi K3 and the Qwen line have set that bar, and Mistral hasn’t committed to meeting it.

The hardware reality check: 1.05T total parameters means roughly 530GB+ of weights at FP4 before KV cache. The 52B active parameters keep per-token compute and bandwidth demands in big-MoE territory rather than dense-1T absurdity, but this is multi-GPU-server or Mac-Studio-cluster hardware, not a single RTX 3090. For what MoE models of this class actually need at home, see the hardware breakdowns at runaihome.com; for the open-weight coding-agent stacks that run on hardware you already own, aifoss.dev’s Continue.dev vs Cline vs Aider comparison is the right starting point.

Verdict: who should actually run Large 4?

Your situation	Run this	What it costs	Where
Daily agentic coding, cost-sensitive	Claude Haiku 5.5	$0.10/$0.50 per MTok (≤100K prompts)	Anthropic pricing
Multi-file work where wrong edits cost review cycles	Claude Sonnet 5.5	$2/$10 per MTok	Anthropic pricing
EU data residency requirement, or 150K+-token sessions you won’t compact, or image input in the loop	Mistral Large 4 via Cline	$0.68/$2.09 until ~Oct 20, then $1.36/$4.18	console.mistral.ai
Open-weight frontier MoE as a coding backend, today	Kimi K3 (DeepSWE 68.0)	see our K3 pricing breakdown	/blog/kimi-k3-cursor-cline-coding-backend-2026/

The two-week launch discount is a fair free-trial window: $15/month-equivalent for a frontier-class generalist with a million-token window is cheap enough to test against your real workload. Just measure it against Haiku 5.5 on the same tasks before the price doubles — on everything we can verify today, the cheap model wins the coding fight.

FAQ

What is the exact API model ID for Mistral Large 4? mistral-large-4 or the pinned mistral-large-4-0 — both are documented API names. The versioned form pins you to release 26.10; the alias will follow future point releases.

Does Mistral Large 4 support function calling for agentic tools like Cline? Yes. Function calling and structured outputs are documented capabilities, and Cline’s native Mistral provider drives its edit loop through them. Tool messages carrying images need middleware (Cline ships it); plain tool calls work out of the box.

Is Mistral Large 4 cheaper than Claude Sonnet 5.5? Yes on list price — $1.36/$4.18 vs $2/$10 per million tokens (and half that during the launch window) — but it also scores far below Sonnet on agentic coding benchmarks (22.7–28.3% vs 70.6% on Terminal-Bench 4). Cheaper per token is not cheaper per completed task if you re-run failures on a stronger model.

When do the Mistral Large 4 open weights come out? Mistral says end of October 2026; reports narrow it to October 27–31. The license is unannounced — treat “open-weight” as unconfirmed until the HuggingFace repo and license text are live.

Sources

• Mistral platform docs (official source repo for docs.mistral.ai) — Large 4 model card, changelog 2026-10-06, pricing schema
• Mistral Large 4 model card — docs.mistral.ai
• Anthropic pricing — Claude Haiku 5.5 / Sonnet 5.5 / Opus 5.5
• Mistral Large 4 Preview — Artificial Analysis intelligence, speed, and verbosity data
• Mistral Large 4 preview: pricing, benchmarks, open weights — DEV Community
• Mistral Large 4 (Le Chonk): Everything We Know — DataCamp
• Cursor community forum — staff confirmation that the OpenAI base-URL override is a supported BYOK route
• Cline repository — native Mistral provider and model catalog

Last updated October 11, 2026. Pricing verified against Mistral’s official docs source and claude.com on this date; the Mistral launch discount is scheduled to end around October 20, 2026. Benchmark scores marked vendor-reported await independent replication.

Know which coding tool is worth paying for

Hands-on comparisons of AI coding assistants and what each one costs to run — including the local-model path. Sent only when something changes. Unsubscribe anytime.

Email address Subscribe