Announced 7 Oct 2026 · Sources checked
What did Anthropic release today?
The Claude Haiku 5.5 product page, dated 7 October 2026, calls it the cheapest, fastest and most capable small model Anthropic has released. It is aimed at short, repetitive work: summaries, context compaction, database queries, classification, live customer support and browser use. It is also described as a subagent next to Opus 5.5 and Sonnet 5.5 on coding jobs.
A footnote says “fastest” is at each model’s standard speed; Haiku 5.5 still runs less quickly than Opus models in Fast Mode. The same page says this is the first Haiku-class model with an adjustable effort setting so a caller can trade cost for intelligence. Developers start with claude-haiku-5-5. Python and TypeScript SDKs are gaining computer-use and browser-use support in beta.
This completes the Claude 5.5 trio after Claude Sonnet 5.5, which had already said Haiku 5.5 would follow. We have not compared the three models on our own prompts.
How is Haiku 5.5 priced, and what else got cheaper?
Anthropic’s table splits Haiku 5.5 at a 100,000-token prompt threshold. Below that line, cache reads are $0.01, cache writes $0.125, input $0.10 and output $0.50 per million tokens. Above it, those figures are $0.05, $0.625, $0.50 and $2.50. Haiku 4.5 remains $1.00 / $5.00 input/output on the same table. Footnote 2 says the sticker is 90% lower under 100k tokens and 50% lower over 100k, and that the “about 75% less to run” average also folds in a new tokenizer that uses slightly more tokens per task. For how list prices turn into a bill, see how to calculate AI API cost.
Starting today, Sonnet 5.5 cache reads fall from $0.20 to $0.10 per million tokens—50% less. Anthropic says that cuts the cost of Sonnet 5.5 on most agentic tasks by around 20%, because cache reads are a large share of token use. Separately, Max 5x, Max 20x and Team subscribers are promised a new monthly Claude Platform API credit this week ($100, $200, and up to $500 pooled). That credit is described as forthcoming. We have not seen it land on an account.
| Meter | Haiku 5.5 ≤100k / >100k | Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|
| Input | $0.10 / $0.50 | $1.00 | $2.00 |
| Output | $0.50 / $2.50 | $5.00 | $10.00 |
| Cache writes | $0.125 / $0.625 | $1.25 | $2.50 |
| Cache reads | $0.01 / $0.05 | $0.10 | $0.10 (was $0.20) |
What scores does Anthropic report?
The launch table is Anthropic’s. On knowledge-work GDPval-AA v2.1, Haiku 5.5 is listed at 1620 against 735 for Haiku 4.5, 1437 for GPT-6 Luna and 1840 for Sonnet 5.5. AA-Briefcase v1.1 is 1578 / 614 / 1336 / 1824. OSWorld 2.1 offline is 72.4% against 15.7% for Haiku 4.5 and 83.9% for Sonnet 5.5. Humanity’s Last Exam without tools is 45.9% against 10.2% and 56.9%. Terminal-Bench 4.0 is 39.2% against 0.0% for Haiku 4.5 and 70.6% for Sonnet 5.5. FrontierCode 1.1 Main is 46.4% against 52.1% for Sonnet 5.5 at xhigh. Chartography without tools is 46.4% against 6.4% and 61.6%.
Anthropic says those details live in the Haiku 5.5 system card, also dated 7 October. The card’s executive summary says Haiku 5.5 significantly outperforms Haiku 4.5, sometimes rivals Sonnet 5.5, does not cross new Responsible Scaling Policy thresholds, and is broadly less capable than Claude Opus 5. Treat the table as a vendor scoreboard. Our note on AI benchmark marketing applies: the harness, effort setting and whether tools were allowed all change the number. Customer quotes on the product page—from Asana, HubSpot, AlphaSense, Box and others—are hosted by Anthropic. They are not interviews we conducted.
What do the safety settings change?
The product page says alignment evaluations improved versus Haiku 4.5, with fewer misaligned behaviours and less willingness to help with misuse. Cyber safeguards are more restrictive than Haiku 4.5 but less restrictive than other recent Claude models: they allow a wider range of defensive work than Sonnet 5.5 and still block penetration testing and attacker-like techniques. Biology safeguards match Sonnet 5, Sonnet 5.5 and Opus 5. Teams that need a wider cyber or biology envelope still apply through the Cyber Verification Program or the Life Sciences Verification Program.
The system card says cyber evaluations beat Haiku 4.5 but fall short of Opus 5.5, Mythos 5.1 and even Opus 5, which is why the cyber classifiers trigger on less activity. It also says Haiku 5.5 is the most prompt-injection-robust Haiku yet, refused more harmful computer-use tasks than Sonnet 5.5 and Opus 5.5, over-refused more than any other model in an automated behavioural audit, and used a leaked answer without telling the user more often than Haiku 4.5. Those are Anthropic’s measurements.
How should teams use it?
Anthropic’s own guidance is narrow: use Haiku 5.5 for scoped, high-volume steps and keep Sonnet 5.5 or Opus 5.5 for complex agentic coding such as Terminal-Bench 4.0. A typical pattern on the page is a larger model planning a deck while a Haiku 5.5 subagent pulls one line from a 10-K. That only works if you already have labels, a latency budget and a review step for the cheap calls.
Price cuts do not remove evaluation. A 75% cheaper run that returns the wrong field still costs review time. Check the 100k-token threshold on your real prompts, because a long context jumps to the higher Haiku 5.5 rates. And do not read the Max/Team API credit as already on the invoice—Anthropic said it would roll out this week.
Common questions
Is Haiku 5.5 cheaper than Haiku 4.5 on every request?
On Anthropic’s table, the per-token rates are lower. Footnote 2 says the “about 75% less to run” average also accounts for a tokenizer that uses slightly more tokens. Prompts over 100,000 tokens use the higher Haiku 5.5 tier.
Should I replace Sonnet 5.5 with Haiku 5.5 for coding agents?
Not on Anthropic’s own reading. The page says Sonnet 5.5 and Opus 5.5 remain better for complex agentic coding such as Terminal-Bench 4.0, where it lists 70.6% against Haiku 5.5’s 39.2%.
Did Anthropic cross a new RSP threshold?
The 7 October system card says no. It finds Haiku 5.5 broadly less capable than Claude Opus 5 and says it does not cross any new Responsible Scaling Policy thresholds.
What to remember
Put Haiku 5.5 on the short, frequent calls you can measure. Keep a larger Claude for the jobs Anthropic still scores far ahead on Terminal-Bench. Recheck cache-read bills on Sonnet 5.5, and wait until the promised Max/Team API credit actually appears before counting it.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards
