The Small Tier Takes the Stage

Some weeks the universe announces itself with a supernova. This one spoke in the quieter language of small things moving fast.

No lab unveiled a new flagship. Instead, the price of an ordinary thought fell again. On 7 October, Anthropic made Claude Haiku 5.5 generally available, from $0.10 / $0.50 per million tokens for prompts up to 100k tokens, with a 1M-token context, and on the same day halved Sonnet 5.5 cache reads from $0.20 to $0.10. A day earlier OpenAI opened a Decisions API in beta that it calls 10x faster than the Responses API. Perplexity cut its decision model's input price from $0.04 to $0.02 per million. In Paris, Mistral put a model with 1 trillion parameters, 52 billion of them active at a time, into public preview.

The wonder is not in any one of these. It is that a routine call to a machine mind now costs a fraction of a cent, and the real work has become deciding which question goes to which mind.

In the News

Anthropic: Haiku 5.5 arrives, and the agents learn to work in phases

On 7 October, Anthropic launched Claude Haiku 5.5 (claude-haiku-5-5), its cheapest and fastest Claude: 1M context, 128k max output, adaptive thinking with effort control, priced from $0.10 / $0.50 per million tokens up to 100k prompt tokens. The same release notes cut Sonnet 5.5 cache reads from $0.20 to $0.10 per million. On 9 October, Claude Managed Agents gained dynamic workflows in beta, described as "a program that runs many agents in phases". The Claude 5.5 family now has its three rungs. Pinned Sonnet 4.5 IDs still retire on 30 November 2026 (model deprecations).

Why it matters. The cheapest Claude now carries the same 1M context as its siblings. High-volume work such as triage, extraction and routing can move down a tier without losing the long-document window.

OpenAI: a Decisions API, Ultrafast for Sol, and a HIPAA path

OpenAI's changelog filled up this week. On 6 October, the Decisions API (v1/decisions) opened in beta on gpt-6-luna. It returns typed yes/no, choice and score answers, and OpenAI calls it "10x faster than the Responses API". Usage tiers were simplified from five to three. On 8 October, OpenAI added Ultrafast mode for GPT-6.1 Sol, "available to all API users, subject to rate limits, with global processing and US and EU data residency". Ultrafast on Astra covered global and US residency only, so EU is new. On 5 October, an in-product flow arrived for accepting the Business Associate Agreement and turning on HIPAA compliance support in API organisation settings. The Agents API is still in beta, and the changelog has no GA entry for it.

Why it matters. EU-resident teams can now use Sol at Ultrafast speed, and teams with narrow yes/no or scoring calls have a purpose-built endpoint. Healthcare builders get a self-serve route to HIPAA support.

Mistral: Large 4 enters public preview

On 6 October, Mistral put Mistral Large 4 into public preview. It is a multimodal mixture-of-experts model with 1T total and 52B active parameters and a 1M context (Mistral release). The changelog says "open weights are coming soon", and Mistral's release points to the end of the month. The weights are promised, not shipped. Launch pricing is 50% off for two weeks. The same changelog retires Mistral-hosted GLM 5.2 on 31 October. Use GLM 5.3 instead.

Why it matters. Europe has a trillion-parameter frontier contender in preview. Until the weights actually ship, treat it as a hosted API, and plan any discount-window testing as if the window could close on short notice.

Elsewhere on the Shelf

Look past the launches and you see the other half of the cosmos: things ending, on schedule, like stars burning through their fuel.

Google's Gemini API changelog shows Nano Banana 2.1 (gemini-nano-banana-2.1) reaching general availability on 6 October. Two days later it retired older siblings. Gemini 3.7 Flash is deprecated and auto-routed to 3.8 Flash, 3.5 Flash routes to 3.6 Flash, and the Deep Research agent deep-research-pro-preview-12-2025 "will be shut down on October 23, 2026". The deprecations page holds two more dates. gemini-omni-flash-preview shuts down on 22 October (move to gemini-omni-1.1-flash), and the shutdown date for antigravity-preview-05-2026, 5 October, has passed (the replacement is -09-2026). Gemini 4 Argon still has no public model ID in the changelog.

Perplexity's changelog carries a hard date of its own. At 00:00 UTC on 24 October, its Agent and Orchestrator APIs retire openai/gpt-5.4, -mini and -nano, plus gpt-5.2, 5.1, 5 and 5-mini. In the same changelog, pplx-decider-v1.1-27b replaces v1 at $0.02 per million input tokens, and the Agent API now supports Haiku 5.5. With OpenAI's Decisions API and AWS's Strands Decider, decision models now form a small class of their own. They are emerging task models, not frontier-LLM peers.

A few names shifted in the starlight. Pages on docs.x.ai are now titled "Release Notes | SpaceXSI Docs" and "Grok Models & Pricing | SpaceXSI Docs", and the X handle now reads @SpaceXSI. DeepSeek's updates page has not moved since 10 September: V4.1-Flash is still the default, and there is still no V4.1-Pro.

Regulators kept their own slower orbit. In the UK, the government accepted all 44 recommendations of the National Commission into the Regulation of AI in Healthcare on 6 October. MHRA AI Airlock Phase 3 is open, with an applicant webinar at 10:00 UK time on 22 October and draft guidance on changes to AI medical devices due by December. In Washington, EO 14434 tells federal agencies to say "Super Intelligence" in non-statutory communications. It is a terminology order, not a new compliance duty, and it leaves regulations, contracts and the statutory AI definition unchanged. In California, Governor Newsom signed a further package of AI bills on 30 September, covering automated decisions in employment, workplace surveillance, AI transparency and health-care AI. In Europe, the Digital Omnibus sets 2 December 2026 for marking AI-generated content from systems already on the market, and moves high-risk obligations to 2 December 2027 and 2 August 2028.

Business Impact

While the labs made thinking cheaper, three organisations published what agents did inside real operations. Each set of figures comes from a vendor's own release.

On 8 October, sportswear company On and Google Cloud reported that AI agents migrated 24 core services, "cutting migration time from three months to just two weeks per service". A small internal team ran 15 migrations in-house with under five minutes of planned downtime per service.

On 6 October, Salesforce published how ENGIE uses Agentforce. It resolves 71% of cases autonomously across billing, invoices and energy use, handles 1,500+ conversations a day at 83% response accuracy, and went live in weeks, compared with nine months for earlier AI projects.

On 12 October, IBM's newsroom described a pilot at Mashreq using IBM Bob for integration engineering. It reported "up to a 60-70% reduction in the time spent on integration design and troubleshooting", with about 4,000 agent requests a month. This is a pilot, not a production rollout.

Three months became two weeks. Nine months became weeks. Hours became minutes. These are company-stated results, not a census. Still, they show where agents are paying back: in the long, careful work of moving, answering and connecting systems.

Takeaway

This was a week for the small tier. Haiku 5.5 brought a 1M-token Claude from $0.10 per million input tokens. Decision endpoints from OpenAI and Perplexity made yes/no answers a product. Sol gained Ultrafast with EU residency. Mistral Large 4 arrived in preview, with its open weights still a promise. Meanwhile the deadline list grew longer than the launch list: 22 October for Omni Flash preview, 23 October for Deep Research preview, 24 October for Perplexity's older OpenAI models, 31 October for Mistral-hosted GLM 5.2, and 30 November for Sonnet 4.5. On, ENGIE and Mashreq showed what agents return when they are given real queues.

The job now is routing: which calls go to the small tier, which go to a decision model, and which still need the flagship. For a facilities team, that might mean a small model answering routine work-order questions while the larger ones handle the harder calls. Put the dates in the calendar before the stars go out.