# Le Chonk Matches Terra on Coding Index: Score 48 [2026] > Source: > Published: 2026-10-31 05:08:37+00:00 Mistral AI’s Le Chonk, formally Mistral Large 4, now sits in the same coding bracket as OpenAI’s GPT-6.1 Terra and Codex. The Paris lab released the model’s open weights on October 29, 2026, and reports from Gigazine and The Decoder say it scored 48 on the Artificial Analysis Coding Agent Index, which puts it level with that closed pair. It arrives with a license that lets most companies run it on their own hardware. This story has several moving parts, and some of them cut against the headline. The model still trails the best Chinese open models on overall intelligence scores, and its license carries a catch for cloud resellers. Mistral Le Chonk coding results are strong, but they are one slice of a broader picture. We covered the [cyber-defense side of this release](https://shattered.io/mistral-large-4-open-weights-cybersecurity-51-percent-2026/) earlier. This analysis focuses on the coding ranking, the token-efficiency claim, the pricing and what it means for teams choosing a model this quarter. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. [Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F) ## What Mistral Announced on October 29 Mistral published the open weights for Mistral Large 4 on October 29, 2026, according to SiliconANGLE and The Decoder. The company nicknames the model Le Chonk, a playful label for a mixture-of-experts system with roughly one trillion total parameters. The Decoder and SiliconANGLE both put the active parameter count at 52 billion per step. The weights sit on Hugging Face, and the model is also reachable through Mistral Studio, the Mistral API, OpenRouter, NVIDIA endpoints and Microsoft Foundry. The release follows a public preview that CNBC dated to October 6. That gap is 23 days. In that window Mistral says it fed preview feedback back into training and kept scaling reinforcement learning. The company reports major gains in coding, defensive cybersecurity and reasoning efficiency. A technical report is listed as coming soon, so some of the numbers below rest on Mistral’s own launch materials and on third-party indexes that relay them. Other published specs include a context window of up to one million tokens, native multimodal input, fluency in more than 160 languages and three reasoning modes (low, medium and high, with high as the default). Gigazine adds that the model takes text and images as input and returns text and code. Note that the spec sheet is unusually broad for an open release. Mistral is pitching Le Chonk as a general-purpose flagship, not a coding specialist. ## How Mistral Le Chonk Scored on the Coding Agent Index The headline number is 48 on the Artificial Analysis Coding Agent Index. Both The Decoder and Gigazine report that score, and Gigazine says it matches GPT-6.1 Terra and Codex. Mistral pairs the model with its Vibe coding tooling in that comparison, per Gigazine’s summary. We could not pull the Artificial Analysis leaderboard directly, so treat the exact rank order as reported rather than independently checked. The index is useful because it tests agentic work, meaning multi-step tasks where a model reads a repository, edits files and runs commands. A plain code-completion test would not capture that. The score of 48 says the model holds up across those loops, and the cost figures later in this piece say it does so cheaply. | Benchmark | Mistral Large 4 result | Reported by | |---|---|---| | Artificial Analysis Coding Agent Index | 48 (on par with GPT-6.1 Terra and Codex) | The Decoder, Gigazine, Mistral | | Terminal-Bench 4.0 | 41% (up from 28% in preview) | The Decoder, SiliconANGLE, Mistral | | FrontierCode 2.0 | 19.7% | The Decoder, Mistral | | Artificial Analysis Cyber Index | 51% | The Decoder, Mistral | | BrowseComp (web browsing) | 84.5% | Mistral | | FinWorkBench | 73.3% | Mistral | | Harmful-request refusal rate | 96.6% | The Decoder, Mistral | The FrontierCode 2.0 figure of 19.7% looks low next to the others, and it deserves a plain reading. The benchmark is built to be hard, so a score under 20% does not signal a weak model. It signals that the test leaves headroom for every vendor. Mistral also cites Surge Code human-preference results in which testers favored Le Chonk over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Gigazine reports a win rate above 50% against Chinese-made models in that evaluation. ## Terminal-Bench 4.0 Jumps From 28% to 41% The sharpest improvement is on Terminal-Bench 4.0, which measures how well a model drives a command line. The preview scored 28%, and the open-weights build scores 41%, a gain of 13 points in under a month. SiliconANGLE says that result ties GLM-5.3, one of the leading Chinese open models. A 13-point jump over 23 days fits Mistral’s account of continued reinforcement learning on the preview. Mistral also says output tokens on [Terminal-Bench](https://shattered.io/claude-opus-5-5-vs-gpt-6-astra-terminal-bench-2026/) fell by 30% compared with the preview. That matters twice. Fewer tokens means lower bills, and it also means faster completions when an agent loops through dozens of steps. The company ties this to its reasoning-effort controls. Developers can pick low, medium or high and trade accuracy against spend on a per-task basis. CNBC’s October 6 report described the preview as the strongest [open-weight](https://shattered.io/deepseek-vs-qwen-vs-mistral-open-weight-models-2026/) model built outside China, while also saying it lagged the frontier in areas such as coding. Three weeks later the coding story reads differently. The index placing it beside Terra suggests the lag shrank fast. Whether it closed fully depends on which benchmark you trust, and the ties reported here come from indexes that favor token-efficient agents. ## The 47% Token Savings Claim, Explained Mistral says Le Chonk uses 47% fewer tokens on the Coding Agent Index than the leading open-weight competitor. SiliconANGLE names that competitor as GLM-5.3. Gigazine’s summary phrases it more loosely as fewer tokens than rivals, so the safest reading is the narrower one. The claim covers one index and one comparison model, not every task and every rival. The arithmetic still matters. If a rival burns 1,000 output tokens on a job, a model using 47% fewer burns 530. At Mistral’s output price of $2.10 per million tokens, that difference compounds quickly across an agent that runs thousands of tasks a day. Token efficiency also cuts latency, because generation time scales with output length. Skeptics will point out that cheaper tokens do not help if a model needs more attempts to succeed. That is a fair worry, and the Coding Agent Index partly answers it because it scores task success alongside cost. Still, every team should rerun its own tasks. A model that wins on a public index can lose on a messy internal monorepo. Parity on an index also hides how teams actually work. A coding agent lives or dies on the boring middle of a task, where it reads a stack trace, picks the right file and avoids breaking a neighboring test. Benchmarks sample that behavior, but they cannot cover every language, build system and legacy quirk. Treat a score of 48 as a strong signal to run a trial, and not as proof that the model will replace your current tool on every repository. ## Pricing: $0.70 In, $2.10 Out, and a $2.06 Task Mistral lists API pricing at $0.70 per million input tokens and $2.10 per million output tokens. Cached input drops to $0.07 per million. The Decoder measured generation speed at 161 tokens per second and an average cost of $2.06 per task on the Artificial Analysis index. The model is also free to download, so teams with spare GPUs can skip per-token fees entirely and pay only for hardware and power. | Spec or price | Mistral Large 4 (Le Chonk) | |---|---| | Input price | $0.70 per million tokens | | Cached input price | $0.07 per million tokens | | Output price | $2.10 per million tokens | | Generation speed | 161 tokens per second | | Average Coding Agent Index task cost | $2.06 | | Context window | Up to 1 million tokens | | Reasoning modes | Low, medium, high (default) | | Total / active parameters | About 1 trillion / 52 billion | The cached-input price is the detail to watch for coding agents. Agents resend large repository contexts over and over, so a 90% discount on repeated input (from $0.70 to $0.07) changes the real cost curve. We have not seen published cached-input rates for the closed rivals in this comparison, so we will not put a number on the gap. Our [earlier report on GPT-6.1 Terra in ChatGPT Work and Codex](https://shattered.io/gpt-6-1-terra-chatgpt-work-codex-api-pricing-2026/) covers how OpenAI packages its side. ## Open Weights, With a License Catch Mistral ships Le Chonk under the Mistral Ecosystem Open-Weight License, which The Decoder and Gigazine both describe as a modified MIT license. Companies can use, modify, fine-tune and commercialize the model. The condition: businesses that run a Model-as-a-Service offering must obtain a separate commercial license from Mistral. That clause targets cloud resellers, not ordinary users. A startup embedding the model in its own product is fine. A host that wants to sell raw API access to the weights has to talk to Mistral first. Gigazine words it as a restriction on third-party API services, which matches. [Open-source](https://shattered.io/best-open-source-llm-october-2026-kimi-glm-deepseek/) purists will note this falls short of a standard MIT or Apache grant, and the license debate will run for weeks. For most buyers the practical effect is small. Enterprises that want to keep source code inside their firewall can download the checkpoint and run it. SiliconANGLE frames this as the main selling point, since sensitive data never leaves the company’s network. Our [comparison of six open-weight model families](https://shattered.io/open-weight-models-license-comparison-6-families-16gb-2026/) lays out how this license stacks up against the others. Open weights also change what a team can do with the model after download. Companies can fine-tune it on private code, pin a specific checkpoint so behavior never shifts under them, and audit how it runs inside their own network. A closed API offers none of that control, and it can change or retire a model on the vendor’s schedule. For regulated teams, version pinning alone can justify a trial, even before price enters the discussion. ## Where Le Chonk Still Trails: The Intelligence Index The coding story has a counterweight. On the Artificial Analysis Intelligence Index, a broad measure across many task types, The Decoder reports that Mistral Large 4 scores 26. That sits below DeepSeek V4.1 Flash and GLM-5.3 at 32 each, and below MiMo-V2.6-Pro at 30. Kimi K3.1 Pro leads that group at 46, though The Decoder notes its weights are not yet public. The outlet’s headline sums it up: a strong coding model that still trails China overall. | Model | Intelligence Index | Gap vs Le Chonk | Note | |---|---|---|---| | Kimi K3.1 Pro | 46 | +20 | Weights not yet public | | DeepSeek V4.1 Flash | 32 | +6 | Chinese open model | | GLM-5.3 | 32 | +6 | Chinese open model | | MiMo-V2.6-Pro | 30 | +4 | Chinese open model | | Mistral Large 4 (Le Chonk) | 26 | 0 | Western open weights | Read the table with care. A six-point gap on a general index does not erase the coding result, because the two indexes measure different things. Mistral built Le Chonk to excel at cyber, coding, manufacturing, finance and multimodal tasks, according to CNBC’s account, and it did not chase a general-knowledge crown. Buyers choosing a coding agent should weight the coding index. Buyers wanting one model for everything should weight the broader score. ## Competitive Comparison: Terra, Codex, GLM-5.3 and Kimi Against the closed pair, Le Chonk wins on openness and, by Mistral’s numbers, on token use. OpenAI’s Terra and Codex keep the advantage of a vendor-managed service with deep product integration. Against the Chinese open models, Le Chonk ties GLM-5.3 on Terminal-Bench, claims a 47% token advantage over it, and wins human-preference tests against several rivals. It loses on the broad Intelligence Index. GLM-5.3 deserves special attention because developers already use it for serious security and patching work. Epoch AI’s test found it could crack six Firefox builds at roughly $1,000 each, which we covered in our [GLM 5.3 Firefox patch-gap report](https://shattered.io/epoch-ai-glm-5-3-firefox-patch-gap-six-versions-1000-each-2026/). That makes it a credible production rival, and a tie with it is a meaningful milestone for a Western lab. ### Gigazine’s Claude Opus 5.5 Detail Gigazine lists Claude Opus 5.5 among the models in the human-evaluation comparison, alongside GLM-5.3 Flash, Kimi K3 and DeepSeek V4.1 Flash. Mistral’s own post names only the Chinese models in its human-preference line. The two accounts differ, and we have no way to settle which framing is right without the full technical report. Until it lands, hold the Opus comparison loosely. ## Market Impact for Closed Vendors and Tool Builders A model at index parity with Terra, priced at $0.70 and $2.10 per million tokens and free to self-host, pressures closed vendors on price. It does not need to beat them. It only needs to be good enough that procurement teams ask why they pay more. Tool builders feel this first, since an editor or agent that supports Le Chonk through OpenRouter can offer customers a cheaper default overnight. The second effect is geographic. Governments and regulated industries in Europe have asked for capable models they can host under local control. SiliconANGLE stresses this sovereignty angle, and Mistral trained the model on Nvidia Grace Blackwell GPUs in European data centers (CNBC says about 4,000 GPUs over two months, SiliconANGLE says 3,800). We use the more conservative description here, which is a few thousand GPUs. For readers tracking the broader field, our [AI and machine learning section](https://shattered.io/ai-machine-learning/) follows each of these launches. Mistral also has the funding to keep pushing. CNBC reports a Series D of 3 billion euros (about $3.4 billion) raised in September at a 21 billion euro valuation. Chief scientist Guillaume Lample told CNBC that capabilities should keep improving as the company scales training capacity after that raise. CEO Arthur Mensch, per SiliconANGLE, framed open source as the way to keep top-tier AI from concentrating in a few hands. ## Historical Context: From Preview to Weights in 23 Days Mistral’s release cadence this month is the real story behind the ranking. The preview arrived October 6 alongside news that Reflection AI had unveiled its own open model the day before, per CNBC, with GLM 5.2-level claims. Open weights followed on October 29. In between, Terminal-Bench scores climbed from 28% to 41% and output tokens fell 30%. Few labs ship that much measurable improvement in under four weeks. Earlier coverage on this site tracked the first wave of results, including our look at the [model’s DeepSWE result](https://shattered.io/mistral-large-4-open-weight-1t-parameters-deepswe-61-7-2026/). The pattern across all of it is consistent. Mistral releases, collects feedback, retrains and re-releases quickly, and each pass moves a coding metric. That loop is the competitive asset, more than any single score. The timeline also explains the cautious CNBC framing. When that article ran, the model lagged the frontier in coding by Mistral’s own admission. The October 29 numbers describe a different checkpoint. Readers comparing old and new coverage should check which build each number refers to, since the preview and the weights release are not the same model. ## What Developers Should Test Before Switching Public indexes are a starting line, not a finish. Pick 20 to 50 real tasks from your own backlog, run them against Le Chonk and your current model, and compare success rate, tokens used and wall-clock time. Log failures by category. That small harness will tell you more than any leaderboard, and it takes an afternoon. ### Start With Reasoning Modes The default mode is high, which spends the most tokens. For routine edits, try medium or low and measure how much accuracy you give up. Mistral built the modes to let teams control spend per task, so a tiered setup often beats a single setting. Use high for hard refactors and low for formatting or boilerplate. ### Decide Between API and Self-Hosting A one-trillion-parameter model is large even with only 52 billion active parameters, because all the weights must sit in memory. Most teams will start on the API or a hosted endpoint, then move to their own cluster if volume justifies it. If you plan to experiment locally with smaller models first, our guide to [local LLM runtimes](https://shattered.io/ollama-vs-lm-studio-vs-llama-cpp-vs-vllm-local-llm-runtimes-2026/) covers the tooling choices. ## Five Predictions for the Next Quarter These are our own expectations, based on the numbers above, and not claims from Mistral or any quoted source. - **Coding tools add Le Chonk as a default option.** With OpenRouter and Microsoft Foundry access already live, integrations take days, not months. - **Rivals respond on price, not benchmarks.** Closed vendors rarely cut headline rates, but expect cheaper cached-input tiers and effort-based discounts. - **The license clause draws pushback.** The Model-as-a-Service condition will spark a debate over what counts as open, and some hosts will negotiate terms. - **Mistral ships another checkpoint before year-end.** The 23-day preview-to-weights cycle and the Series D cash make a quick follow-up likely. - **Independent evaluations narrow the claims.** Once the technical report lands, third parties will test the token-savings and parity figures, and some will shrink. ## The Takeaway on Mistral Le Chonk Coding Results Le Chonk earns its place in the top coding tier on the evidence published so far. A score of 48 on the Coding Agent Index, a 41% Terminal-Bench result and a claimed 47% token saving add up to a competitive coding model with open weights and low prices. The caveats are real too. It trails Chinese open models on the broader Intelligence Index, the license restricts resellers, and the full technical report is still pending. For engineering teams, the sensible move is a measured trial, not a migration. Run your own tasks, watch the token counts and keep a fallback. For the industry, the message is clearer. A European lab matched the closed leaders on a coding index within weeks of a preview, and it handed the weights to everyone. ## FAQ: Mistral Le Chonk and Coding Rankings **What is Le Chonk?** Le Chonk is the nickname for Mistral Large 4, a multimodal mixture-of-experts model with roughly one trillion total parameters and 52 billion active. Mistral released its open weights on October 29, 2026. **Does Le Chonk really match GPT-6.1 Terra on coding?** Gigazine and The Decoder report a score of 48 on the Artificial Analysis Coding Agent Index, which Gigazine describes as on par with GPT-6.1 Terra and Codex. We could not check the live leaderboard ourselves, so the parity claim is as reported. **Is Mistral Large 4 free to use?** The weights are free to download under the Mistral Ecosystem Open-Weight License, a modified MIT license. Businesses that operate a Model-as-a-Service offering need a separate commercial license. The hosted API costs $0.70 per million input tokens and $2.10 per million output tokens. **How does it compare with Chinese open models?** It ties GLM-5.3 on Terminal-Bench 4.0 at 41%, and Mistral claims 47% fewer tokens on the Coding Agent Index. On the broader Intelligence Index it scores 26 against 32 for GLM-5.3 and DeepSeek V4.1 Flash, per The Decoder. **Where can developers access it?** Mistral lists Mistral Studio, its API, OpenRouter, Hugging Face, NVIDIA endpoints and Microsoft Foundry. **What reasoning modes does it offer?** Low, medium and high, with high as the default. Lower modes use fewer tokens and cost less per task. **Has the technical report been published?** Not yet. Mistral lists it as coming soon, so independent verification of the efficiency and parity claims is still pending.