Dateline: October 9, 2026 | Next update: October 16, 2026
After a fortnight of frontier launches, this week went the other way. The notable model is the cheapest one in the lineup, and the notable features are about limits rather than reach: what an agent's commands may touch, which hosts it may call, which URLs it may fetch. Anthropic tightened all three in Managed Agents, and GitHub made a local sandbox generally available at no extra cost. OpenAI, meanwhile, shipped an endpoint that deliberately does less — the Decisions API answers from a fixed set rather than generating prose, and is faster for exactly that reason.
Claude / Anthropic
★ Claude Haiku 5.5, and the price cliff at 100,000 tokens
Anthropic's fastest model gains the million-token context window the rest of the lineup already had, with 128K maximum output and adaptive thinking steered by the effort parameter, defaulting to medium. It is built for classification, extraction, routing and subagent work — high volume, latency sensitive.
The pricing is tiered, and the step is steep. Up to a 100,000-token prompt: $0.10 input, $0.50 output per million. Above 100,000 tokens: $0.50 input, $2.50 output — five times the rate, on the whole request. A classification job that usually sends 80,000 tokens and occasionally sends 120,000 does not get 50% more expensive on those calls. It gets five times more expensive. If you are batching documents into one prompt, the batch size is now a pricing decision.
Model ID: `claude-haiku-5-5`; `anthropic.claude-haiku-5-5` on Bedrock | Context: 1M tokens | Max output: 128K, up to 300K on the Batches API with `output-300k-2026-03-24` | Pricing up to 100K-token prompts: $0.10 input / $0.50 output; cache read $0.01; 5m cache write $0.125; 1h cache write $0.20 | Pricing over 100K-token prompts: $0.50 input / $2.50 output; cache read $0.05; 5m cache write $0.625; 1h cache write $1 | Batch API: 50% off | Thinking: adaptive, default effort `medium` | Knowledge cutoff: June 2026 | Retirement: not sooner than October 7, 2027
Best for: High-volume classification and routing, provided you keep prompts under 100,000 tokens. Instrument prompt length before you migrate, not after.
⚠ Haiku 5.5 counts tokens differently, and breaks five things
Haiku 5.5 uses the newer tokenizer introduced with Claude 4.7, so the same text counts as roughly 30% more tokens than it did on Haiku 4.5.
Thirty percent more tokens for identical input is a real cost change that no price list will show you, and it interacts with the 100,000-token cliff above: a prompt that measured 85,000 tokens on Haiku 4.5 lands around 110,000 on Haiku 5.5 and crosses into the five-times tier. Five code changes also return 400 errors: manual thinking with budget_tokens, temperature, top_p, top_k and assistant prefill. The computer_20250124 tool is rejected — use computer_toolset_20260801 on the Claude API and Google Cloud, or computer_20251124 on Bedrock. Replaying thinking blocks after changing system, tools or earlier turns can also 400, and structured outputs are not available on Bedrock for this model.
Tokenizer: same as Claude 4.7 and later; identical text is approximately 30% more tokens than on Haiku 4.5 | Returns 400: manual thinking (`budget_tokens`), `temperature`, `top_p`, `top_k`, assistant prefill | Computer use: `computer_20250124` returns 400; use `computer_toolset_20260801` on the Claude API and Google Cloud, `computer_20251124` on Bedrock | Thinking blocks: replaying them after changes to `system`, `tools` or earlier turns can return 400; blocks are account-bound | Structured outputs: not available on Bedrock for Haiku 5.5 | Thinking display: adaptive thinking is on by default, thinking text is omitted unless `thinking.display` is `"summarized"`, and it counts toward output tokens
Best for: Read the migration guide before changing the model string. The tokenizer change alone justifies re-measuring a sample of real prompts.
★ Sonnet 5.5 cache reads halve, and paid plans gain API credits
Prompt cache reads on Sonnet 5.5 dropped from $0.20 to $0.10 per million tokens — from 10% of base input to 5%, matching the rate Opus 5.5 already had. Cache writes and every other price are unchanged. Separately, Claude Max and Team plans now include monthly API credits.
Halving the cache read rate is a targeted cut: it only helps workloads that reuse a large prompt prefix across many calls. If you run an agent with a long system prompt and a stable tool set, your largest single line item just fell by half with no code change. If every request is different, nothing changed. The plan credits matter for a different reason — a team on Max that wanted to prototype against the API previously had to open a separate billing relationship to do it.
Sonnet 5.5 cache read: $0.20 → $0.10 per MTok, now 0.05× base input (previously 0.1×) | Unchanged: $2 input, $10 output, 5m cache write $2.50, 1h cache write $4 | Claude Max and Team plans: now include monthly API credits | Opus 5.5 cache read: already $0.20 per MTok at 5% of its $4 base input
Best for: Anyone running a long, stable system prompt at volume. Check what share of your Sonnet spend is cache reads — that share just halved.
★ Managed Agents get dynamic workflows, and much tighter networking
Two changes in the same product, pulling in opposite directions. On October 9, dynamic workflows entered beta: an agent can write a program that runs many agents in phases and combines their results, with the server executing it as a workflow run and reporting progress through workflow_run.* events. Enable it by setting the agent's multiagent field to {"type": "multiagent_20261001", "workflows": {"type": "enabled"}}.
Two days earlier, the networking rules around those agents got considerably stricter. Under `limited` networking, `allowed_hosts` now also governs `web_search` and `web_fetch`, and a disallowed fetch returns `url_not_allowed`. An `allowed_domains` entry outside `allowed_hosts` now returns a 400 at session creation rather than failing later. And `web_fetch` will only fetch URLs already present in the session — anything else returns `url_not_in_prior_context`.
Dynamic workflows (October 9, beta): beta header `managed-agents-2026-04-01`; enable with `"multiagent": {"type": "multiagent_20261001", "workflows": {"type": "enabled"}}`; track via `workflow_run.*` events on the session event stream | Networking (October 7): with `limited` networking, `allowed_hosts` governs `web_search` and `web_fetch`; disallowed fetches return `url_not_allowed` | `allowed_domains` outside `allowed_hosts` returns 400 on session create or update | `allowed_hosts` matches exact hosts unless an entry starts with `*.` | `web_fetch` only fetches URLs already in the session; others return `url_not_in_prior_context`
Best for: The networking change is the one to action. If your agent fetches URLs it discovers mid-task, that pattern now fails by design — check before it surprises you in production.
★ Claude Code — hooks that can block, and effort on subagents
Haiku 5.5 became the default Haiku model on the Anthropic API in 2.1.293. Two other additions are worth knowing.
In 2.1.295, hooks gained onFailure: "block": a hook that cannot start, times out, or exits with an unexpected code now blocks the action rather than letting it through. If you have been using a hook as a policy gate, it was previously fail-open — a broken hook meant no gate at all. That is now a choice rather than the only behaviour. In 2.1.292, the Agent tool gained an effort parameter, so a subagent can run at a different thinking depth from the main session.
2.1.293 (October 7): Claude Haiku 5.5 added as the default Haiku on the Anthropic API; `agentType` added to the `subagentStatusLine` payload | 2.1.295 (October 8): `onFailure: "block"` for command and HTTP hooks; Program Status Protocol (OSC 7501) so terminals can show whether Claude Code is working, waiting or done; warning when `claude plugin install`/`enable`/`disable` writes to a settings file that does not load | 2.1.292 (October 6): `effort` parameter on the Agent tool; `--marketplace
Best for: Anyone enforcing policy through hooks. Fail-open was a quiet weakness; `onFailure: "block"` closes it.
Plans and pricing
Two cuts and no rises. Haiku 5.5 enters below the model it replaces for prompts under 100,000 tokens, and Sonnet 5.5 cache reads halved.
Haiku 5.5: $0.10/$0.50 per MTok up to 100K-token prompts; $0.50/$2.50 above | Haiku 4.5: $1/$5 per MTok, flat | Sonnet 5.5: $2/$10 per MTok, cache read now $0.10 (was $0.20) | Opus 5.5: $4/$20 per MTok, cache read $0.20 | Fable 5.1: $10/$50 per MTok | Batch API: 50% off input and output on all models | Claude Max and Team: now include monthly API credits | No retirements announced this period
Best for: Re-run the arithmetic on anything currently on Haiku 4.5. Below 100K tokens the new model is an order of magnitude cheaper; above it, the gap closes sharply.
ChatGPT / OpenAI
★ The Decisions API — an endpoint that deliberately does less
Announced in limited preview at DevDay, the Decisions API is now in beta. It takes text and images and returns a typed answer from a set the developer defines in advance — classify this, route that, pick the next action. OpenAI claims results ten times faster than the Responses API.
The speed claim follows from the constraint rather than from a faster model. A model that must choose from a fixed list is not generating prose and does not need to be parsed, retried or validated against a schema afterwards. For the classification and routing steps buried inside most automations — is this invoice or a receipt, does this ticket go to billing or support — that removes a whole layer of glue code. It is also the same job Claude Haiku 5.5 was built for, which makes this week an unusually direct comparison.
Endpoint: `v1/decisions` | Model: `gpt-6-luna` | Input: text and images | Output: typed answers from a developer-defined finite set | Use cases named by OpenAI: classify content, route requests, choose an agent's next action | Claimed performance: 10× faster than the Responses API | Status: beta, following a limited preview announced at DevDay on September 29
Best for: The classifier sitting in the middle of an existing workflow, currently implemented as a chat call plus a JSON parser plus a retry.
★ Five usage tiers become three
OpenAI collapsed its five API usage tiers into three — Build, Launch and Grow. Organisations move up automatically as total credit purchases reach each tier's minimum.
Fewer tiers means fewer rate-limit thresholds to plan around, which is simpler. It also means the gaps between them are wider, so the jump from one tier's limits to the next is larger than it used to be. If your scaling plan was built around reaching tier 4 by a particular date, that milestone no longer exists and the replacement milestone sits somewhere else.
Previous structure: five usage tiers | New structure: three — Build, Launch, Grow | Progression: automatic as total credit purchases reach each tier's minimum | OpenAI published no pricing figures with the change
Best for: Anyone whose capacity planning referenced the old tier numbers. Re-read your current limits rather than assuming they mapped across.
★ Ultrafast comes to GPT-6.1 Sol, with EU data residency
Ultrafast mode, which shortens the gap between generated output tokens, now works with GPT-6.1 Sol on the Responses API. It is open to all API users subject to rate limits.
The detail that changes who can use it is data residency. When Ultrafast launched with GPT-6 Astra at DevDay it was US-only, which ruled it out for a lot of European work. The Sol version offers global processing with both US and EU data residency — so the speed tier is now available to teams that could not touch it ten days ago.
Model: `gpt-6.1-sol` | Parameter: `service_tier: "ultrafast"` | API: Responses API | Effect: reduces the time between output tokens | Availability: all API users, subject to rate limits | Processing: global, with US and EU data residency | Compare: GPT-6 Astra Ultrafast, announced September 29, is US data residency only
Best for: European teams who read the Ultrafast announcement at DevDay and stopped at the residency line. It no longer applies to this model.
★ HIPAA support, and a self-service BAA
Eligible administrators can now accept OpenAI's standard Business Associate Agreement and enable HIPAA compliance support from within organisation settings.
Location: API Organization settings → General | Mechanism: eligible admins accept the standard Business Associate Agreement (BAA) in product | Result: HIPAA compliance support enabled for the organisation | Also this week: the `chat-latest` snapshot now points to the latest model in ChatGPT for Plus, Pro, Business and Enterprise users — OpenAI recommends the GPT-6 model family for production API use rather than `chat-latest`
Best for: Healthcare teams who had a signed-BAA requirement blocking a pilot. Note the warning on `chat-latest`: a snapshot that moves is not what you pin production to.
Plans and pricing
No rate changes were published. The commercial change is structural — the tier system was simplified rather than repriced.
Usage tiers: five reduced to three (Build, Launch, Grow), effective October 6, 2026; automatic progression on total credit purchases | Decisions API: beta, pricing published on the pricing page rather than in the changelog | Ultrafast for GPT-6.1 Sol: pricing linked from the pricing page, no figures given in the release note | No published price changes to GPT-6.1 Sol or GPT-6 Astra between October 3 and 9, 2026
Gemini / Google
★ Nano Banana 2.1 reaches general availability
Google's image model left preview. gemini-nano-banana-2.1 updates gemini-3.1-flash-image with better visual quality, closer prompt adherence, more consistent characters across multiple turns, and improved text rendering inside images.
It also adds extreme aspect ratios — 1:4, 4:1, 1:8 and 8:1 — at 1K, 2K and 4K. Those are banner and tower shapes rather than illustration shapes, which points at where Google expects this to be used: generated assets for interfaces and advertising, where the ratio is dictated by a slot rather than by composition. Multi-turn character consistency is the other practical gain, because it is what makes a sequence of images usable as a set rather than as five unrelated pictures.
Model: `gemini-nano-banana-2.1`, generally available | Replaces: `gemini-3.1-flash-image`, now deprecated with no shutdown date announced | Improvements: visual quality, prompt adherence, multi-turn character consistency, text rendering | New aspect ratios: 1:4, 4:1, 1:8, 8:1 at 1K, 2K and 4K | Pricing: none published with the release
Best for: Teams generating interface or campaign assets at fixed ratios. Migrate off `gemini-3.1-flash-image` on your schedule — no shutdown date is set, but deprecated models do not improve.
⚠ Three deprecations, one with a shutdown date two weeks out
Google deprecated two Flash models and one Deep Research agent on the same day, and only one carries a date.
The two Flash deprecations auto-route, so nothing breaks: `gemini-3.7-flash` requests go to `gemini-3.8-flash`, and `gemini-3.5-flash` requests go to `gemini-3.6-flash`. Google asks you to update the model string anyway. The Deep Research agent is the one that needs action — `deep-research-pro-preview-12-2025` shuts down on October 23, two weeks after the notice, and migration means changing the `agent` parameter in `interactions.create` to `deep-research-preview-04-2026` or `deep-research-max-preview-04-2026`.
`gemini-3.7-flash` → `gemini-3.8-flash`, requests auto-routed | `gemini-3.5-flash` → `gemini-3.6-flash`, requests auto-routed; `gemini-3.8-flash` also listed as an alternative | `deep-research-pro-preview-12-2025`: shuts down October 23, 2026; migrate to `deep-research-preview-04-2026` or `deep-research-max-preview-04-2026` by changing the `agent` parameter in `interactions.create` | `gemini-3.1-flash-image` (October 6): deprecated, no shutdown date
Best for: Two weeks is short notice for a shutdown. If you use the Deep Research agent, that change goes in this sprint.
Plans and pricing
No pricing changes were published for the Gemini API in this window.
No published Gemini API price changes between October 3 and 9, 2026 | Nano Banana 2.1: general availability, no pricing published with the release | Deprecations announced: `gemini-3.7-flash`, `gemini-3.5-flash`, `gemini-3.1-flash-image`, and the `deep-research-pro-preview-12-2025` agent
Microsoft Copilot
★ Local sandboxing reaches general availability, at no extra cost
Commands and tools that Copilot starts on a developer's own machine now run inside a sandbox with restricted access to files, network, credentials and other system capabilities. It is built on Microsoft eXecution Container, which maps one sandbox policy onto each operating system's native controls.
Two things make this more significant than a feature release. First, it is included with Copilot at no additional cost — a containment boundary shipped as a default rather than as an enterprise upsell, which is not how this capability has usually been packaged. Second, enterprise-managed settings can require sandboxing and enforce policies that a developer cannot weaken, which turns it from a personal safety net into something an organisation can actually attest to. Policies apply to tool execution regardless of which model Copilot is using.
Engine: Microsoft eXecution Container (MXC), mapping a common sandbox policy to native OS controls | Surfaces: GitHub Copilot CLI, the GitHub Copilot app, and VS Code sessions using Agent Host | Platforms: Windows, macOS, Linux | Controls: which files and directories commands may read or modify; internet and local network access; access to Git and GitHub CLI credentials; sandboxing of local tools and services such as local MCP and language servers where supported | Scope: applies to tool execution regardless of model | Cost: included with GitHub Copilot at no additional charge | Enterprise: managed settings can require sandboxing and enforce policies developers cannot weaken
Best for: Anyone who has been running an agent against a repository and quietly hoping. This is the control that makes that defensible, and it costs nothing.
★ A purpose-built model for finding leaked secrets
GitHub trained a dedicated model for secret detection. Rather than matching known token formats, it reads the surrounding code to identify likely credentials — including passwords that have no recognisable pattern at all. It does not generate code or prose.
Format matching has always had an obvious blind spot: a password in a config file looks like any other string. A model that reads context can flag it, which is the category of leak that survives every regex-based scanner. The rollout is staged rather than uniform — AI-detected alerts have already switched to the new model for existing customers, GHES 3.23 alerts are in public preview, AI push protection is in private preview, and the Copilot security review checks are listed as coming soon. Push protection and security review usage consumes AI Credits, and security review is off by default.
Model: GitHub fine-tuned, purpose-built for secret detection; reads surrounding code; does not generate code or prose | Secret scanning alerts: existing customers with AI-detected Password alerts automatically moved to the new model; included with GitHub Secret Protection and GitHub Advanced Security at no extra charge, including GHES 3.23 (public preview) | Push protection: checks unstructured credentials at push time so a secret can be removed before entering history; GitHub Enterprise Cloud or GitHub Team with GHSP or GHAS; not in the GHES release; private preview; consumes AI Credits | Copilot `/security-review`: secret classifier checks alongside the existing LLM-based review; coming soon in private preview; Copilot Pro, Pro+, Max, Free, Student, Business and Enterprise without a GHSP or GHAS licence; consumes AI Credits; off by default | No accuracy or performance figures published
Best for: Repositories where the real risk is a password in a config file rather than a well-formed API token. Note the AI Credits consumption before enabling it broadly.
★ Haiku 5.5 in Copilot, and local model discovery in the CLI
Claude Haiku 5.5 appeared in Copilot on October 7, the day Anthropic released it — the third consecutive same-day model addition. The Copilot CLI separately gained the ability to discover local models.
Claude Haiku 5.5 in GitHub Copilot: October 7, 2026, same day as the Anthropic release | Copilot CLI: can now discover local models | Also this week: Copilot code review gained new organization billing options and controls (October 8); an IDE update is required to restore agent activity in Copilot usage metrics (October 6); weekly release post covering October 5 published October 9
Best for: Teams who want a cheap model for the routine passes and a frontier model for the hard ones, chosen per task in the same tool.
Plans and pricing
No seat price changes. The commercial movement is in billing controls and in what consumes AI Credits.
No Copilot seat price changes announced between October 3 and 9, 2026 | Local sandboxing: included at no additional cost | Copilot code review: new organization billing options and controls, October 8 | AI Credits consumed by: AI push protection, and Copilot security review secret classifier checks | Secret scanning AI alerts: no extra charge with GitHub Secret Protection or GitHub Advanced Security | Model added: Claude Haiku 5.5, October 7
Best for: Check which of the new security features draw on AI Credits before switching them on across an organisation — two of the three do.