Mistral Large 4, Haiku 5.5 and DeepSeek V4.1 Flash: The AI Price Floor Moves

The week in five lines

• Mistral Large 4 (October 6): public preview of a 1-trillion-parameter multimodal model, API listed at $1.36 input and $4.18 output per million tokens, weights promised by the end of October.
• Claude Haiku 5.5 (October 7): $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, 90% below Haiku 4.5. Sonnet 5.5 cache reads halved the same day.
• DeepSeek V4.1 Flash (released September 10, back in focus October 7): MIT-licensed weights, native image input, $0.15 input and $0.60 output off-peak.
• GPT-6 in ChatGPT (October 7): answers built from a fixed library of interface components, reaching free users on GPT-6 Luna.
• OpenAI's math repository (October 7): three manuscripts withdrawn after a sign error, 300 of 719 top-line results now formally verified.

Three price lists, one direction

Within two days, Mistral opened a preview of its largest open-weight model, Anthropic cut the price of its small model by 90%, and OpenAI put GPT-6 in front of ChatGPT's free users. A month earlier, DeepSeek had shipped a multimodal Flash model in the same price band as the new Haiku. The same week, OpenAI withdrew three mathematics papers after a sign error was found in one of them.

For teams that ship, the common thread is tier selection. The question has moved from "which frontier model" to "which workload needs one at all". Below: what each vendor published, what compares on paper, and what is still a vendor claim.

Mistral Large 4: an open-weight preview with a date attached

On October 6, Mistral launched a public preview of Mistral Large 4. According to its announcement, the model is natively multimodal, has 1 trillion parameters with 52 billion active, and was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, which also serve the preview. Mistral says a significant share of the training data covered more than 160 languages, including every official EU language. The preview API is live in Mistral Studio at $1.36 per million input tokens and $4.18 per million output tokens on the model's listing. The weights are promised "by the end of the month", so you cannot self-host it today.

Every benchmark number comes from Mistral: 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4, and a blind human evaluation with Surge AI in which the preview ranked second of five models on coding quality (3.74 out of 5), behind Claude Opus 5 (4.22) and ahead of Kimi K3 and two GLM versions. Treat that as a vendor-run comparison until independent evaluators reproduce it.

The cybersecurity pitch is the most specific part. On a test that asks a model to reproduce a real vulnerability in open-source software and then patch it, Mistral reports 82% and says Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the task. That does not make ML4 a model without refusals: Mistral also reports that ML4's average refusal rate on malicious cyber prompts (JailbreakBench, StrongREJECT, AgentHarm) is higher than any other open-weight model it tested. The pitch is a different refusal boundary, one that allows proving a flaw is real while still declining clearly malicious requests. Until the weights ship, cybersecurity firms, vetted partners and state authorities are red-teaming a version with reduced moderation and expanded cyber capabilities, so the preview currently has two access levels.

The practical constraint is arithmetic. One trillion parameters at 8 bits is roughly a terabyte of weights, so "open weights" here means a multi-GPU cluster or a hosted deployment, not a workstation. For most teams the realistic gain is a European-operated API priced below Sonnet 5.5's $2 / $10 list, plus the option to move the same model in-house later.

Haiku 5.5 and DeepSeek V4.1 Flash: the cheap tier gets crowded

On October 7, Anthropic released Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output for prompts up to 100,000 tokens, against $1.00 and $5.00 for Haiku 4.5. Above 100,000 tokens the price rises to $0.50 and $2.50. Anthropic says the shorter range covers about 90% of requests to the previous Haiku and, accounting for token usage, estimates Haiku 5.5 costs around 75% less to run on average. It also halved Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens, which it estimates cuts the cost of most agentic work on Sonnet by about 20%.

Haiku 5.5 is the first Haiku with an adjustable effort setting, so the cost-versus-quality dial Sonnet and Opus already expose now exists at the cheapest tier. Anthropic positions it for summaries, classification and subagent work next to Opus 5.5 and Sonnet 5.5, and states that the two larger models "remain better choices for complex agentic coding tasks".

DeepSeek released V4.1 Flash on September 10, with weights on Hugging Face under an MIT license and native image input. It came back into the conversation through an October 7 post from a developer who has used it heavily for a month and says that mid-session they cannot tell it apart from Opus. The Hacker News thread passed 1,100 points and 980 comments. The author calls it subjective and still brings in Opus 5.5 for a final review on critical tasks, so read it as an adoption signal, not a benchmark.

List prices per million tokens, from each vendor's page:

Sources: Anthropic, DeepSeek pricing, Mistral.

Two readings follow. On cache misses and output, Haiku 5.5 is cheaper than DeepSeek's Flash at list price, so an open-weight model from China is no longer the default budget option. On cache hits DeepSeek is cheaper at both peak and off-peak rates, and that is the line item that dominates long agent sessions. DeepSeek's peak window (01:00 to 04:00 and 06:00 to 10:00 UTC, weekdays, excluding Chinese public holidays) doubles every rate, which matters if batch jobs run on a European morning.

We would not choose between them on price alone. A cheaper token only saves money if the model needs no extra retries or longer outputs to reach the same answer. The useful move is to rerun your own eval set on both and compare cost per correct answer, using your real cache hit rate rather than the vendor's best case.

OpenAI: Intelligent UI in the free tier, and three withdrawn proofs

On October 7, OpenAI announced GPT-6 for the Chat experience in ChatGPT, with a feature called Intelligent UI. The model composes answers from a library of native, streamable components (charts, forms, buttons, small tools), and a compiler renders the interface while the answer is still being generated. It rolled out that day to Plus, Pro, Business and Enterprise on GPT-6 Sol, and from the next day to Free and Go on GPT-6 Luna. OpenAI also says the model can interleave reasoning with answering, and that GPT-6 Instant starts answering 44% sooner than GPT-5.6 Instant on questions that need web search. Those are internal measurements. The models behind Work and Codex are unchanged.

For builders, the design is the interesting part. A fixed component library plus a streaming compiler is a constrained alternative to letting a model emit arbitrary HTML. If you generate UI from model output in your own product, that constraint keeps the output reviewable and the rendering predictable.

The same day, OpenAI's math repository logged a withdrawal of three manuscripts. A sign error in "Algebraicity of Weil classes on split abelian eightfolds" invalidates a stabilization-trace cancellation argument, and two papers that relied on its construction were withdrawn with it. The entry lists fourteen other manuscripts repaired and says 300 of 719 top-line results, about 42%, are now formalized. The page does not say how the original proofs were produced, so we draw no conclusion about model capability.

The lesson is about process. The error survived until after publication, and formal verification is the only check in that log that runs mechanically. An eval suite works the same way: a result nobody has checked mechanically is still a claim.

What we're watching next week

First, whether Mistral ships the Large 4 weights by the end of October, and whether independent evaluators confirm the coding and cybersecurity numbers once anyone can run the model. Second, we plan to run the same set of production-style prompts through Haiku 5.5 and DeepSeek V4.1 Flash with caching enabled, because list prices alone will not settle the choice. If the cost-per-correct-answer numbers are worth reading, we will publish them with the method.

If you want that comparison run on your own prompts before you switch tiers, get in touch.