Haiku 5.5 pricing by tier
USD per million tokens. “Over” is prompts above 100,000 input tokens.
| Line | ≤ 100K | > 100K |
|---|---|---|
| Input | $0.10 | $0.50 |
| Output | $0.50 | $2.50 |
| 5-minute cache write | $0.125 | $0.625 |
| 1-hour cache write | $0.20 | $1.00 |
| Cache read | $0.01 | $0.05 |
Cost & pricing
Claude Haiku 5.5 and the Decisions API: the October cheap-tier reset.
In the week after the September 28–29 workhorse reset, both labs re-plumbed the cheapest tier. Anthropic launched Claude Haiku 5.5 at $0.10 / $0.50 per million tokens on October 7 — with a tokenizer that makes the same text about 30% more tokens than Claude Haiku 4.5 — and halved Claude Sonnet 5.5 cache reads to $0.10. OpenAI added Ultrafast mode to GPT-6.1 Sol, now with EU data residency, and shipped the Decisions API, a /v1/decisions surface that bills input tokens only.
- Claude Haiku 5.5
- $0.10 / $0.50
- Sonnet 5.5 cache read
- $0.20 → $0.10
- Decisions API
- input only
per 1M tokens · 1M context · released Oct 7, 2026
halved Oct 7; base input unchanged at $2
$0.10 / 1M in · no output, cache-read or cache-write charge
The Claude 5.5 family, and where the new floor sits
USD per million tokens, standard tier, from Anthropic’s Claude Haiku 5.5 page observed October 12, 2026. Haiku is priced in two tiers: up to 100,000 input tokens, and above.
| Model | Price / MTok | Context | Max output | Latency | Default effort |
|---|---|---|---|---|---|
| Claude Fable 5.1 claude-fable-5-1 | $10 / $50 | 1M | 128K | Slower | high |
| Claude Opus 5.5 claude-opus-5-5 | $4 / $20 | 1M | 128K | Moderate | medium |
| Claude Sonnet 5.5 claude-sonnet-5-5 | $2 / $10 | 1M | 128K | Fast | high |
| Claude Haiku 5.5 claude-haiku-5-5 | From $0.10 / $0.50 | 1M | 128K | Fastest | medium |
Sources: platform.claude.com · Claude Haiku 5.5 · anthropic.com · Introducing Claude Haiku 5.5 for the launch, family table and pricing tiers.
Claude Haiku 5.5: a cheaper headline and a heavier tokenizer
The launch price is $0.10 / $0.50, but two multipliers decide the real bill: the tokenizer and the 100K threshold.
The same text, about 30% more tokens
Claude Haiku 5.5 uses the same newer tokenizer as Claude 4.7 and later, so the same text counts as roughly 30% more tokens than on Claude Haiku 4.5. A workload budgeted on token counts captured from Haiku 4.5 bills more tokens at the same text volume, which can offset the lower headline rate — and a prompt that crosses the 100K boundary steps to a rate five times the base. It ships on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry, with retirement not sooner than October 7, 2027; adaptive thinking is on by default at medium effort, and thinking text is omitted unless thinking.display is set to summarized.
What breaks migrating Haiku 4.5 → Haiku 5.5
Each item is a hard 400 error or a per-platform capability gap, not a warning.
- Manual extended thinking via budget_tokens returns a 400 — control depth with the effort parameter instead.
- A non-default temperature, top_p or top_k returns a 400, as on Claude Sonnet 5.5.
- Assistant message prefill returns a 400.
- computer_20250124 returns a 400: use computer_toolset_20260801 on the Claude API and Google Cloud, or computer_20251124 on Amazon Bedrock.
- Sending thinking blocks back after a change to system, tools or earlier turns returns a 400.
- On Amazon Bedrock, structured outputs are not available for Claude Haiku 5.5.
Sources: platform.claude.com · Release notes (breaking changes, Bedrock structured-output gap) · platform.claude.com · Claude Haiku 5.5 (sampling and thinking contract).
Sonnet 5.5 cache reads were cut in half on October 7
One number from last week’s report is now stale.
Anthropic lowered the prompt cache-read price on Claude Sonnet 5.5 from $0.20 to $0.10 per million tokens — a 0.05× multiplier on the $2 base input price instead of 0.1×. Cache writes and all other prices are unchanged. The same 0.05× cache-read multiplier now applies to Claude Opus 5.5, while Claude Fable 5.1 and Claude Mythos 5.1 read cache at 0.025×. A prompt-cache cost model built on the launch rate overstates the cached leg by 2×; the October 5 report in this series has been re-baselined.
GPT-6.1 Sol Ultrafast: 6× standard, and now EU-eligible
USD per million tokens as input / cached input / cache write / output, from OpenAI’s pricing page observed October 12, 2026. Ultrafast has separate rate limits from Standard and Fast modes.
| Model | Standard (short) | Ultrafast (short) | Ultrafast (long >272K) | Ultrafast residency |
|---|---|---|---|---|
| gpt-6.1-sol | $2.00 / $0.10 / $2.50 / $10.00 | $12.00 / $0.60 / $15.00 / $60.00 | $24.00 / $1.20 / $30.00 / $90.00 | US, EU and global |
| gpt-6-astra | $10.00 / $1.00 / $12.50 / $50.00 | $60.00 / $6.00 / $75.00 / $300.00 | $120.00 / $12.00 / $150.00 / $450.00 | US and global only |
Ultrafast mode is now available to all API users on both GPT-6 Astra and GPT-6.1 Sol, with the same 6× premium over standard on Sol ($12.00 / $60.00 short context). The material difference is residency: GPT-6.1 Sol Ultrafast supports US and EU data residency and global processing, whereas GPT-6 Astra Ultrafast supports US data residency and global processing only — so EU-residency traffic can now buy the latency tier on Sol, which it could not on Astra. Default Ultrafast limits are per the new three-tier structure (Build 1M / Launch 4M / Grow 40M TPM on Sol). OpenAI recommends WebSockets for agentic workloads, since network overhead erodes the latency gain on plain HTTP.
OpenAI’s Decisions API: a new route with an input-only meter
Released in public beta on October 6, 2026. It is not a Responses or Chat Completions route.
What it is
- Evaluates text, images or both and returns typed answers about 10× faster than the Responses API, on the dedicated POST /v1/decisions endpoint.
- Three question types: predicate (a 0–1 probability), choice (one of your supplied values), and score (a probability-weighted average over ordered levels).
- gpt-6-luna is the only model supported at beta. Use Structured Outputs with the Responses API when you need your own JSON schema, and function calling when the model must request a tool.
Why the meter matters
- On gpt-6-luna, a Decisions request costs $0.10 per million input tokens, and that is the whole bill: there are no cache-read, cache-write or output-token charges.
- Regional processing premiums and long-context input multipliers still apply. Other requests on gpt-6-luna keep the standard model rates.
- The surface supports Zero Data Retention and HIPAA for eligible customers, with data residency in the United States and Europe (EEA + Switzerland) — the same regional split as the model, stated per endpoint.
Also official in the October 5–10 window
Three changes that move a routing or lifecycle assumption without shipping a flagship.
Google: two Flash models now auto-route, silently
On October 8, Google deprecated gemini-3.7-flash (replacement gemini-3.8-flash) and gemini-3.5-flash (replacement gemini-3.6-flash), with all requests to the deprecated IDs automatically routed to the replacement rather than failing — no shutdown date is announced. Separately, the deep-research-pro-preview-12-2025 agent shuts down on October 23, 2026; migrate the agent parameter to deep-research-preview-04-2026 or deep-research-max-preview-04-2026. On October 6, gemini-nano-banana-2.1 reached GA and gemini-3.1-flash-image was deprecated with no announced date.
OpenAI: gpt-5.1 and the classic TTS models now have shutdown dates
Announced October 1, 2026 with six months’ notice: gpt-5.1 and gpt-5.3-codex are removed April 1, 2027 (replacement gpt-6-sol) and gpt-5.4-nano the same day (replacement gpt-6-luna). The text-to-speech models tts-1, tts-1-hd and the gpt-4o-mini-tts snapshots are removed January 6, 2027 in favour of gpt-realtime-2.1-mini. Usage tiers were simplified from five to three — Build, Launch and Grow — on October 6.
What changes for a routing or margin decision
- Re-measure tokens, not just price, before migrating to Haiku 5.5. The newer tokenizer inflates the same text by roughly 30% versus Haiku 4.5, so a cost model built on Haiku 4.5 counts understates token volume — and any prompt that crosses 100K input tokens steps to 5× the base rate. Compare cost per request on the new tokenizer, not cost per token.
- A Haiku 4.5 → 5.5 swap is a contract migration. Sampling parameters, assistant prefill, manual budget_tokens thinking and the older computer-use tool each return a 400. Validate the exact request shape before wiring the fallback.
- Bedrock is not at parity for Haiku 5.5. Structured outputs are unavailable there, and the computer-use tool ID differs by platform. A route that relies on structured output cannot simply switch provider for this model.
- Update prompt-cache models on Sonnet 5.5 now. The cached leg is $0.10, not $0.20; a strategy that assumed “cache reads are 10% of input” overstates the cached portion by 2×. Cache writes ($2.50 for 5 minutes) remain a separate purchase.
- Decisions is a new route on a new meter. A gateway that proxies Responses and Chat Completions has no /v1/decisions path, and a token budget written for output-heavy calls will overstate a Decisions request, which bills input tokens only. Route classification and routing work here, and keep schema generation on the Responses API.
Sources
All official provider pages observed October 12, 2026 (Europe/Paris). Prices are USD per million tokens as the provider states them.
- [1]https://platform.claude.com/docs/en/release-notes/overview
- [2]https://platform.claude.com/docs/en/models/haiku-5-5/overview
- [3]https://www.anthropic.com/claude-haiku-5-5
- [4]https://platform.claude.com/docs/en/models/sonnet-5-5/overview
- [5]https://platform.claude.com/docs/en/about-claude/pricing
- [6]https://platform.claude.com/docs/en/about-claude/model-deprecations
- [7]https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html
- [8]https://developers.openai.com/api/docs/changelog
- [9]https://developers.openai.com/api/docs/guides/decisions
- [10]https://developers.openai.com/api/docs/guides/ultrafast-mode
- [11]https://developers.openai.com/api/docs/pricing
- [12]https://developers.openai.com/api/docs/deprecations
- [13]https://ai.google.dev/gemini-api/docs/changelog
- [14]https://ai.google.dev/gemini-api/docs/deprecations
Related live views: the September 28–29 workhorse reset, the non-token meter report and the OpenAI vs Azure AI deployment comparison render the current catalog records.
Deployment intelligence, explained
Why the cheap tier decides more routes than the flagship
Flagship releases set the ceiling; the small, cheap tier carries the volume. Claude Haiku 5.5 and OpenAI's Decisions API both target high-volume, latency-sensitive work — classification, routing, extraction, subagents — which is where most tokens in a production application are actually spent. A price or contract change here moves the default route for a whole workload class, not just the top-end option.
A cheaper headline price is not a cheaper bill. Claude Haiku 5.5 uses the same newer tokenizer as Claude 4.7 and later, so the same text counts as about 30% more tokens than on Claude Haiku 4.5, and a prompt over 100,000 tokens moves to a rate five times the base rate. Read the token count your workload produces on the new tokenizer, and the rate tier it lands in, before assuming a migration lowers cost.