Mistral’s Le Chonk is now a top open model for coding, ranking alongside Codex GPT-6.1 Terra

Mistral released the weights for Mistral Large 4 (ML4), a trillion-parameter open model, and reported a 41% score on Terminal-Bench 4.0, up from 28% in preview and one point ahead of GLM-5.3; its overall Index score was 48 versus GLM-5.3’s 49. Mistral said ML4 uses 47% fewer tokens per task on average than comparable models and scored just above DeepSeek V4.1 Flash on a defensive-cybersecurity benchmark. The model was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months, according to Mistral.

Analysis

The investable implication is a shift in where AI economics accrue, not proof that Mistral is taking material revenue from listed labs. Open-weight models can move value from proprietary per-token pricing toward deployment, orchestration, security, and compute; cheaper workloads may expand usage even as unit prices compress. That creates a two-sided effect for NVIDIA: broader inference and sovereign deployments support accelerator demand, while better capability per unit of compute weakens the assumption that every workload requires frontier-scale clusters. Mistral’s use of NVIDIA hardware is directionally supportive, but a single training run does not establish incremental demand or utilization.

For Anthropic (ANT), the risk is medium-term pricing and differentiation pressure if enterprises can self-host credible coding and defensive-security models. The counterweight is that benchmark leadership and vendor-reported token savings do not establish production reliability, total cost of ownership, support quality, or willingness to accept open-model security and governance burdens. Mistral is private, so there is no direct equity expression.

Near term, expect limited fundamental repricing absent customer adoption or pricing evidence. Over 1–3 months, watch independent evaluations, enterprise deployment announcements, and any changes in API pricing or retention. Over 6–18 months, the structural question is whether open models become good enough for routine workloads, leaving proprietary labs to monetize the highest-stakes tasks. The contrarian risk is treating efficiency as purely bearish for compute: lower cost can induce substantially more usage, but that must show up in utilization or guidance. Thesis weakens if independent tests fail to reproduce performance or if enterprise adoption remains confined to pilots.

Key Decisions for Investors

• No immediate directional trade: the announcement is a competitive signal, not verified evidence of revenue displacement or a change in accelerator demand.
• Keep ANT on a competition watchlist; reassess only with evidence of enterprise workload migration, price concessions, or weaker retention/guidance. A move in ANT based solely on benchmark headlines is vulnerable to reversal.
• For NVDA, treat this as a modest positive for the breadth of the open-model ecosystem, not a standalone buy catalyst. Confirm against data-center demand, utilization, and management commentary on inference before adding exposure.
• Track independent benchmark replication, production deployments, and total cost per completed task over the next 1–3 months. Failure to reproduce vendor-reported efficiency or security results would weaken the open-model displacement thesis.