Claude Haiku 5.5 Ships From $0.10/$0.50 per 1M Tokens
Every new model this week entered our directory as a conditional pick for the same reason: the numbers that sell it are the vendor's own. Claude Haiku 5.5 is now Anthropic's cheapest current model, Mistral Large 4 is a priced public preview whose weights are still a promise, and Kolibri-1 ships Apache 2.0 weights that need about 78 GB of GPU memory.
68 announcements scanned · 17 mattered · 0 verdicts changed
Act before next issue
Re-test before moving work from Haiku 4.5 to Haiku 5.5
Claude Haiku 5.5 lists at $0.10/$0.50 per 1M tokens for prompts up to 100K and $0.50/$2.50 above that. Anthropic's model page says its tokenizer counts about 30% more tokens for the same text, and any non-default temperature, top_p or top_k returns a 400 error. We rate it conditional.
Details & stepsSpend any Claude Code cloud-session credit before 4 November
BleepingComputer reports that Anthropic gave individual Pro subscribers $100 and Max subscribers $250 in credits for Claude Code cloud sessions only. The claim window closed on 7 October and claimed balance expires on 4 November. We found no Anthropic page stating the offer.
Details & stepsTrial Mistral Large 4, but do not commit to it yet
Mistral opened Large 4 as a public preview on 6 October at $1.36 input and $4.18 output per 1M tokens. Weights are promised by the end of October, no licence appears on the pages we read, and most benchmarks are Mistral's own. We rate it conditional.
Details & stepsOn our radar
Kolibri-1: Aleph Alpha's English and German open-weight model
Released 3 October with Apache 2.0 weights on Hugging Face. The FP8 weights need about 78 GB of GPU memory, the pages we read name no hosted API or price, and every benchmark comes from Aleph Alpha's own harnesses. We rate it conditional.
In review, not yet published
None of this week's 17 news drafts has passed our source review, so we report no figures or quotes from them yet: the Claude Haiku 5.5 launch and Sonnet 5.5 cache-read change, Mistral Large 4, Anthropic's three-tier Cyber Verification Program, Claude Code 2.1.290 to 2.1.295 hook fixes, OpenAI's Decisions API beta and usage tiers, GPT-6 as ChatGPT's default, Anthropic's 2026 usage policy, Anthropic's Cyber Mission, Wikimedia's report on OpenAI agents, OpenAI's EU text watermark, Claude Docs, Slides and Design leaving beta, Nano Banana 2.1 and the Veo 3.1 preview shutdowns, Kolibri, ElevenAgents' default voice model, an LMCache flaw, Meta Muse on iPad and EmbeddingGemma 2. Each is written up when it publishes.
Reflection Beam and Step 5 Preview
Beam is early access only, with weights promised later this month and Reflection's own benchmarks. Step 5 Preview has an OpenRouter listing but no StepFun-dated release. Both held until a reader can download or buy them from a primary source.
Google's universal Gemini agent and Anthropic's Claude Startups offer
The Gemini Enterprise agent is a private preview for enterprise customers. The Claude Startups expansion was read only as a headline. Held until GA with a price, or until the Anthropic post is read.
Decision models: Strands Decider 2B and Musubi PolicyLM-1.7B
AWS's Strands Decider primary is now read, but whether small decision models belong in our model directory is not yet settled. Musubi's model is feed-only. Both held until that scope question is ruled.
Docker Agent and Whistle
Docker's YAML-defined multi-agent builder is a trending repo with no dated release read. Whistle, a 16.9 MB on-device speech-to-text model, has vendor-only benchmarks. Held until a release note or an independent measurement.
Held on a headline or feed summary only
Not yet read, so none of these is written up: Meta and Microsoft reportedly reducing employee use of Claude, USA Today's copyright suit against OpenAI, an AI agent fleet on Tencent infrastructure targeting Amap, the ARTEX pentesting tool used against South Korean financial firms, OX Security's scan of 15,465 MCP servers, a Claude diary entry reported to police, Google AI Edge Foresight, the SynthID Detector site, Strata's single-GPU Qwen claim and a Tech Times MCP exposure headline.
OpenAI personnel and research stories
Fired OpenAI safety researchers disputing misconduct claims, OpenAI's mathematics release and its withdrawal of three papers, and fake AI ad portals phishing credentials. None changes a product, access path or reader setting yet.
Open verdict watches, unmoved
Google's deprecations page still says Gemini 2.5 Pro is not deprecated, with no shutdown date. Gemini 4 Argon is still rolling out soon. Grok Bot and DeepSeek V4 Pro show nothing new. Cursor's OpenAI help page could not be read by any fetcher, so its rating stays conditional. Each entry keeps its current rating.
Filtered out
- A verdict claim that Claude Haiku 5.5 should cost GPT-6 Luna its rating. Luna is already conditional rather than a recommended pick, and the only evidence is Anthropic's own table about a rival, so nothing flips.
- A GLM-5.3-Flash pricing post, held as a duplicate of our 14 September report on the end of the launch discount.
- OpenAI feed items: customer case studies, partnerships, an ads format, research posts and a GPT-6 model guide. No tool, model or price change.
- Anthropic's $150M Genesis Mission commitment and Claude Blog case studies and guides. No product change.
- Funding and consumer news, among them Arena at $3.1B, Manus at $500M, Nous at $1.5B and Lambda at $4B. None moves a tool, model or price.
- Non-AI security items, including an Atlassian CVE, SonicWall, FortiBleed, npm malware and a ccTLD certificate hijack. Not AI tools, or not a reader setup.
- Re-reports of stories already in our pipeline: Anthropic's usage policy, Mistral Large 4, Reflection Beam, Haiku 5.5 and Claude Docs.
- Research, opinion and off-beat items, including about 300 off-beat Hacker News stories, demos and essays, and off-beat items from The Verge and TechCrunch.
Our take: Our verdict on this week: Claude Haiku 5.5 is the model to test first, and none of the week's three new models earns more than conditional yet. Haiku 5.5 lists at $0.10 input and $0.50 output per 1M tokens for prompts up to 100K, against $1/$5 for Haiku 4.5. Prompts over 100K cost $0.50/$2.50, Anthropic's model page says the new tokenizer counts about 30% more tokens for the same text, and every benchmark is Anthropic's own. Mistral Large 4 is a public preview at $1.36/$4.18 per 1M tokens, with weights promised by the end of October and no licence on the pages we read. Kolibri-1 is Aleph Alpha's English and German model with Apache 2.0 weights, which need about 78 GB of GPU memory in FP8. The directory also added a DeepSeek Harness vs OpenCode comparison. The one news post we published this week covers a reported Claude Code cloud-session credit that expires on 4 November. The week's 17 news drafts, including the Haiku 5.5 launch post, Anthropic's three-tier Cyber Verification Program and OpenAI's Decisions API beta, are still in review, so their figures are not in this issue. No rating moved on a published post. How we count: one daily scan on October 9 covered six days from October 3, because no scan ran October 4 to 8. It logged 68 source lines. Grouped into stories, that is 5 act, 1 verdict claim, 12 fyi and 21 watch stories, plus 17 noise lines. 17 stories became drafts after the verdict claim was rejected, 1 news post was published, and 0 verdicts changed.
Neomanex Utility editorial
Get issue 28
Next week's verdict changes and deadlines, in your inbox.