Recent AI news · Original report: 2026-10-08 (KST)

Mistral Large 4 preview, Claude Haiku 5.5 pricing, cyber verification tiers (Oct 8, 2026)

This article was compiled with AI. This English translation preserves the original report's collection date, findings, and limitations. Please consult the linked sources for current information. Originally collected: 2026-10-08 10:15 (KST, UTC+09:00).

Translated into English: 2026-10-08 (KST). The translation date does not indicate a new source check.

At a glance

This week's three announcements come down to who can use what. Mistral promised model weights anyone can download, Anthropic opened its cheapest model to the free plan, and cyber security capabilities went the other way, with access released step by step to vetted professionals only. All scores are figures reported by the companies themselves.

This week's announcements

We opened the official posts ourselves, up to 10:15 a.m. Korea time on October 8, 2026. Dates are the ones printed on the original posts.

Mistral Large 4: preview now, weights at the end of the month

France's Mistral opened a public preview of Mistral Large 4 on October 6. For now it is available only through the API on Mistral Studio, and the company said it will release the weight files by the end of the month.

• Size: The announcement says 1 trillion parameters, and the model documentation says 1.05 trillion. The active parameters that actually run at any one time number 52 billion, and the model reads text and images together.
• Training: It was trained on 3,800 NVIDIA GPUs in Mistral's own data centers in Europe, on data spanning more than 160 languages, including every official language of the European Union.
• Price: The model documentation lists $1.36 per million input tokens and $4.18 per million output tokens, and in the same place shows half those amounts, $0.68 and $2.09, as a "sale price."
• Scores the company cited: 61.7% on DeepSWE v1.1, which measures long software engineering tasks. In security tests, it scored 82% on a task that reproduces and patches real vulnerabilities, the highest of any model, and solved 93% of Cybench, a set of 40 exercises drawn from security competitions.
• Before the weights ship: The company said it is giving the same model with reduced moderation to security firms and state authorities, among others, for real-world testing.

Claude Haiku 5.5: the small model got much cheaper

On October 7, Anthropic released Claude Haiku 5.5, its smallest and fastest model. The company wrote that it costs about 75% less to run on average than the previous model, Haiku 4.5.

• Price: For requests up to 100,000 tokens, it is $0.10 per million input tokens and $0.50 per million output tokens. Longer requests over 100,000 tokens cost $0.50 for input and $2.50 for output. Haiku 4.5 was $1 for input and $5 for output.
• Who can use it: It can be selected on every claude.ai plan including the free plan, and it is on the API and the AWS, Google Cloud, and Microsoft clouds. It can read up to 1 million tokens at once.
• Scores the company cited: On OSWorld 2.1, which has the model operate a computer screen directly, it scored 72.4%, up sharply from Haiku 4.5's 15.7%. On Terminal-Bench 4.0, which measures complex coding, it scored 39.2%, a wide gap from the 70.6% of the larger Sonnet 5.5. The company itself wrote that Haiku suits narrow work like summarization and classification.
• What changed alongside it: The cache read price for Sonnet 5.5 dropped from $0.20 to $0.10 per million tokens. Max subscribers receive $100 to $200 in API credits each month.

Cyber security access: three tiers

On October 6, Anthropic said it had reorganized its Cyber Verification Program, which opens cyber capabilities to security professionals, into three tiers. It explained that its generally available models block most cyber work.

• Defense Access: Defending systems you own. Security teams at companies, universities, and government bodies, operators of infrastructure such as regional hospitals, and individual researchers with a track record. The company said it will respond within a few days.
• Red Team Access: Adds authorized penetration testing. Only organizations can apply, and review takes a few weeks. Actions such as deploying ransomware are still blocked in real time at this tier.
• Specialized Access: Only vetted organizations that test systems touching human life or markets, such as power grids and telecom networks. Reviews are conducted together with the US government.
• Test results the company cited: Across 50 attempts at attack scenarios, every attempt was blocked outside the program, while in the Red Team tier the model completed 34 with no blocks.

What it means and for whom (our estimate)

• For families using AI at home, the closest change is that Haiku 5.5 is open on the free plan. But the company itself wrote that larger models are better for complex work, so tasks that need long reasoning are probably still better left on the default model.
• Mistral's weight release is news that widens the path for organizations to run a large model on their own servers, and it is unlikely to be something ordinary users feel right away.
• Two companies saying on the same day that "stronger capabilities go only to vetted security professionals" reads as a signal that AI cyber capabilities are splitting into general release and expert access.

What we could not confirm

• Not tested by us: We have not run any of the three ourselves.
• Mistral's sale price: The documentation does not say how long the half-price period lasts or on what terms. The weight release date and license are only given as "end of the month" and "Open."
• Haiku 5.5 in Korean: The announcement gives no per-language figures. The 75% saving is the company's own calculation, and a footnote says the new tokenizer uses slightly more tokens for the same work.
• Other companies: OpenAI's October 7 post about GPT-6 would not load when we checked, so we did not cover it.

Sources

• Mistral, Introducing Mistral Large 4 (2026-10-06) https://mistral.ai/news/mistral-large-4/
• Mistral documentation, Mistral Large 4 model page https://docs.mistral.ai/models/mistral-large-4-0
• Anthropic, Introducing Claude Haiku 5.5 (2026-10-07) https://www.anthropic.com/claude-haiku-5-5
• Anthropic, Claude Haiku product page https://www.anthropic.com/claude/haiku
• Claude Platform documentation, release notes https://platform.claude.com/docs/en/release-notes/overview
• Claude Platform documentation, models overview https://platform.claude.com/docs/en/about-claude/models/overview
• Anthropic, Expanding the Cyber Verification Program (2026-10-06) https://www.anthropic.com/news/cyber-verification-program