Mistral halves ML4 token pricing to undercut GLM-5.3 on cost and sovereignty

Mistral's decision to release the open weights for Mistral Large 4 on October 29 was the headline, but the sharper move was buried in the price sheet. The Vector Wire terminal's pricing history shows ML4's API rate moved from $1.36/$4.18 per million tokens (input/output) on October 21 to $0.70/$2.10 at the current listing, a reduction of roughly 49% on input and 50% on output. That positions a trillion-parameter frontier model at half the token cost of its closest open-weight rival, GLM-5.3, which lists at $1.4/$4.4 per million tokens as of November 1. The price cut also tracks a broader API commoditization trend: GPT-6.1 Sol dropped its output price from $10 to $2.50 between September and November, and GPT-6 Luna now lists at $0.10/$0.125. But Mistral's move is sharpened by a sovereignty pitch that no Chinese or American lab can replicate. The thesis: Mistral is using price plus European provenance as a wedge to pry enterprise and government buyers away from Chinese open-weight models, in a market where API pricing alone is no longer a durable moat.

The pricing history tells the story in three steps. ML4 entered preview on October 6⁶. By October 21, the API rate stood at $1.36/$4.18. With the open-weights release on October 29, Mistral dropped to $0.70/$2.10. GLM-5.3 lists at $1.4/$4.4 as of November 1. The net effect is that ML4's general-availability price undercuts every comparable open-weight competitor, while the model itself improved substantially during three weeks of additional reinforcement learning^10,11. Note the comparison to Mistral's own prior generation: ML3 lists at $0.50/$1.50, so ML4 costs only 40% more on output for a model that jumped 52 Elo points on Arena Vision (1219.94 to 1272.17, independently verified).

The performance case rests on a narrow but real set of gains. On Terminal-Bench 4.0, ML4's score climbed from 28% in the October 6 preview to 41% at release, a 13-percentage-point jump¹. On the Artificial Analysis Coding Agent Index, ML4 scored 48, one point behind GLM-5.3's 49 but level with Codex GPT-6.1 Terra's 48. Mistral claims ML4 uses 47% fewer tokens per task on average than GLM-5.3. That efficiency claim is vendor-stated and unverified. If it holds, the effective cost gap widens beyond the list-price difference; if independent benchmarks show token efficiency is comparable to GLM-5.3, the cost advantage narrows to the list-price gap alone. Either way, ML4's list price is half GLM-5.3's, and the self-hosting option eliminates per-token fees entirely.

Mistral trained ML4 on 3,800 Nvidia Grace Blackwell GPUs in its European data centers⁴. The model uses a mixture-of-experts architecture with 1.05 trillion total parameters but only about 52 billion active per token, which keeps inference costs lower than the headline parameter count would suggest. It supports a one-million-token context window, fluency in over 160 languages, and three reasoning-effort settings¹⁵. The weights went live on Hugging Face under a modified MIT-style license permitting commercial use¹³.

The sovereignty argument is where Mistral differentiates from both closed Western labs and Chinese open-weight competitors. "Companies need to be able to control where those models get deployed," Mistral Chief Revenue Officer Marjorie Janiewicz said. "They need to own those models so that the day somebody decides to unplug them, they have business continuity"³. CEO Arthur Mensch framed the broader objective: "AI should be a tool for empowerment, not dominance"². For buyers who need data residency and supply-chain independence, the combination of open weights, European-hosted training infrastructure, and competitive pricing creates a procurement story that GLM-5.3 cannot easily replicate regardless of its benchmark scores. But sovereignty cuts both ways: the same open weights that guarantee business continuity also guarantee that adversaries can download the model, a tension the cybersecurity evidence below makes explicit.

The cybersecurity angle is the sharpest illustration of that tradeoff. Mistral says ML4 topped open-weight scores on Cybench and scored 51% on the Artificial Analysis Cyber Index⁸. On the B3 Agent Security Benchmark, resilience to adversarial attacks rose from 93.3% in preview to 97.1% at release. But Mistral itself acknowledged the risk our prior coverage flagged: "Though we observed the model refusing to comply with black-box attacks against live targets, we see a potential for uplift to conduct real attacks," the company wrote in its release blog post^[1]. Pierre Stock, Mistral's VP of Science, told CSO that subsequent safeguard updates would be applied to the managed API and future releases, not to copies already downloaded. Once the weights are in the wild, Mistral loses control over how they are used, and the company's own risk assessment makes that tradeoff explicit rather than hidden. Recent research underscores the broader pattern: a systematic evaluation of adversarial jailbreaking attacks across seven LLMs, including Mistral-7B, found persistent vulnerabilities in open-weight architectures. The realistic buyer profile is not a signals-intelligence agency but an enterprise that needs European data residency, auditability, and freedom from vendor lock-in, and can tolerate the dual-use risk because its use case does not intersect with the offensive-capability concern.

The pricing move lands in a market where the cost of frontier intelligence is falling fast. Anthropic's Claude Opus 5.5 lists at $4/$20 per million tokens. OpenAI's GPT-6.1 Sol sits at $2/$2.50. ML4 at $0.70/$2.10 is a fraction of those closed-model prices, and it comes with the option to skip the API entirely and self-host. For organizations that can provision the hardware (the model can run on as few as four Nvidia B200 or B300 GPUs), the marginal cost drops further.

Cloudera's partnership, announced at its Evolve26 conference, illustrates the enterprise channel Mistral is building. The collaboration pairs Cloudera's data platform with Mistral's models, with joint engineering teams helping customers choose models and infrastructure for specific use cases⁵. Cloudera's senior director of product marketing, Wim Stoop, described the strategy as bringing "AI to where the data is and where the data lives". That on-premises pitch aligns precisely with the sovereignty and cost arguments Mistral is making: if the model is free to download and the data never leaves your network, the total cost of ownership shifts from per-token API fees to hardware and engineering labor.

ANALYSIS Aggressive pricing on ML4 is easier to sustain with substantial capital backing (Mistral raised $3.5 billion in a Series D round), but it also raises the question of how Mistral monetizes a model whose weights are free. If GLM-5.3 simply matches ML4's pricing, the API price wedge vanishes, and Mistral's differentiation narrows to sovereignty and the European training footprint. The company is competing for enterprise spend against OpenAI (which raised $30 billion in October 2026) and Anthropic ($63 billion in debt financing, October 2026). The API at $0.70/$2.10 serves customers who want managed infrastructure; the open weights serve those who want full control and pay nothing for the model itself. For context, Anthropic's Claude Code tool alone runs at a $2.5 billion annual rate; Mistral has disclosed no comparable revenue figure.

Pierre Stock called ML4 "merely a stepping stone," telling reporters that Mistral wants "to be in a position in which we dictate the frontier". That ambition requires sustained investment in post-training, the phase where ML4's gains were concentrated. The 13-percentage-point jump on Terminal-Bench 4.0 between preview and release came from reinforcement learning, not additional pre-training.

Mistral's claim that ML4 uses 47% fewer tokens per task than GLM-5.3 is the load-bearing number in the cost argument, and it needs third-party validation. Whether GLM-5.3's pricing moves in response also matters: as of November 1, GLM-5.3's list price is double ML4's for comparable coding performance (GLM-5.3 scores 49 vs. ML4's 48 on the Artificial Analysis Coding Agent Index), a gap that invites a countermove. DeepSeek's IPO in July signals that Chinese open-weight labs have their own capital access for a price war. Enterprise adoption signals from Cloudera and other channel partners over the next quarter will test whether the sovereignty pitch converts to revenue. And whether Mistral's acknowledged "potential for uplift to conduct real attacks" triggers regulatory scrutiny under the EU's evolving AI framework, particularly given that downloaded weights cannot be retroactively patched, will shape the viability of the open-weight model for every lab, not just Mistral.