Anthropic's Claude Haiku 5.5 Is Its Cheapest Small Model Yet

Anthropic released Claude Haiku 5.5 on Wednesday, October 7, 2026, describing it as the company's cheapest, fastest and most capable small model yet. The model, which carries the ID claude-haiku-5-5, is the third in the Claude 5.5 family in the past month, after Sonnet 5.5 and Opus 5.5. Reuters reported that the release pace is widening Anthropic's lineup as the company prepares for a planned initial public offering. Anthropic is still private, and the IPO has not happened. (Anthropic, Reuters)

Related reading: Anthropic's Claude Duet Works in the Background, and Asks Before It Acts · Anthropic Cuts Live Internet Access for Its Internal Evaluations After Its Agents Broke the Rules.

The pitch is aimed at volume. Anthropic lists high-volume, cost-sensitive jobs as the target: summaries, conversation compaction, database queries, and classification or routing. The company also positions the model as a subagent beneath Sonnet 5.5 and Opus 5.5 on coding work, and as the choice for speed-sensitive tasks such as live customer support and browser use.

A context window of that size changes what the cheap calls can carry. Classification and routing decisions can run against a large document or a long conversation history in a single pass at the low input rate, rather than being chunked across many calls. For coding work, a Haiku subagent can hold a large codebase in context while the larger model plans the task, which is the arrangement Anthropic describes when it places Haiku 5.5 beneath Opus 5.5 and Sonnet 5.5.

What Claude Haiku 5.5 costs and carries

Anthropic published two price tiers. For prompts up to 100,000 tokens, which it says covers about 90% of prior Haiku requests, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens; cache reads are $0.01 and cache writes $0.125. Above 100,000 tokens the price rises to $0.50 input and $2.50 output, with cache reads at $0.05 and cache writes at $0.625. The model has a 1-million-token context window and supports up to 128,000 output tokens. Anthropic says it costs about 75% less to run on average than Haiku 4.5, with the saving at 90% for requests up to 100,000 tokens and 50% above that. The company credits part of the difference to a new tokenizer that uses slightly more tokens per task, which trims the headline saving.

The two tiers are the part developers will watch. The higher band above 100,000 tokens costs five times as much on input and five times as much on output, which is the price of putting a very large prompt into one call. Cache reads at $0.01 per million tokens are the other lever: an agent that reuses the same instructions or document across many calls pays a fraction of the input price each time it does.

The benchmarks are Anthropic's own

Anthropic published a set of comparisons against Haiku 4.5, Sonnet 5.5 and OpenAI's GPT-6 Luna. On computer use (OSWorld 2.1, offline subset) Haiku 5.5 scored 72.4%, against 15.7% for Haiku 4.5, 48.9% for GPT-6 Luna and 83.9% for Sonnet 5.5. On Humanity's Last Exam, a multidisciplinary reasoning test, it scored 45.9% without tools and 57.4% with them, up from 10.2% and 18.7% for Haiku 4.5. On agentic coding (Terminal-Bench 4.0) it reached 39.2%, against 0.0% for Haiku 4.5, 16.4% for GPT-6 Luna and 70.6% for Sonnet 5.5. It also posted GDPval-AA v2.1 of 1620, versus 735 for Haiku 4.5, plus FrontierCode 1.1 of 46.4% and Chartography of 46.4%.

Every one of those figures is Anthropic's own and should be read as a vendor claim, not an independent measurement. The pattern is the same across the set: Haiku 5.5 sits far ahead of the previous Haiku and below Sonnet 5.5, which is the ordering Anthropic itself describes. The company stresses that Sonnet 5.5 and Opus 5.5 remain the better choice for complex agentic coding. The independent benchmark index Vals AI lists Haiku 5.5 among the models it tracks, but the launch comparisons were run and published by Anthropic. (Vals AI)

Computer use is the benchmark that most directly matches the jobs Anthropic is selling Haiku 5.5 for. The jump to 72.4% from Haiku 4.5's 15.7% is the widest gap in the set, and it lands the model above GPT-6 Luna's 48.9% on the same offline subset while leaving it below Sonnet 5.5's 83.9%. Anthropic pairs that result with browser use as one of the jobs Haiku 5.5 is meant to take, so the computer-use number is the one that carries the most weight in the pitch. It is also the one that most needs an outside check, since only Anthropic has run it so far.

Haiku 5.5 is built for volume, and the benchmarks Anthropic published show it beating its predecessor by wide margins while staying below the larger models in the same family.

The cybersecurity safeguards sit between two tiers

Anthropic says Haiku 5.5 shows major alignment improvements over Haiku 4.5, and it drew a specific line on cybersecurity. The safeguards are more restrictive than Haiku 4.5's but less restrictive than Sonnet 5.5's: the model permits a wider range of defensive work but blocks penetration testing and techniques more likely to be used by attackers. Reuters reported it as Anthropic's first Haiku with built-in safeguards for a narrow set of high-risk cybersecurity requests, with most everyday tasks unaffected. Biology safeguards match those of Sonnet 5, Sonnet 5.5 and Opus 5.

The tiering shows how labs set safety policy by model lineup rather than by one setting for everything. Placing the small model's guardrails between the last Haiku and the larger Sonnet lets Anthropic widen what cheap runs can do for defenders while keeping the highest-risk requests inside the tier it watches more closely. That is the company's stated logic; the launch material does not include an outside evaluation of the safeguards.

The comparison Reuters drew, a first for the Haiku tier, points to a wider shift in how these guardrails are written. A model that blocks penetration testing but still allows defensive security work is separating the same technique used by an attacker from its use by a defender. That is a judgment about intent, not one about the request itself, and Anthropic says it has tuned the model to make it for a narrow set of high-risk requests while leaving everyday tasks alone. Nothing in the release explains how the model decides where that line falls.

The small-model tier is where the fight moved

Haiku 5.5 lands in the segment that matters most to developers building agents: small, fast models that get called many times per task. A cheap model that is good enough at routing, summarizing and tool use lets a developer fan out a single request into dozens of subagent calls without the bill scaling the same way. Anthropic points Haiku 5.5 at that role beside Opus 5.5 and Sonnet 5.5.

Anthropic paired the launch with two pricing moves the same day. It halved Sonnet 5.5 cache reads to $0.10 per million tokens from $0.20, which it says makes Sonnet 5.5 about 20% cheaper on most agentic work. It also added a monthly API credit for Max and Team subscribers: $100 a month for Max 5x, $200 a month for Max 20x, and up to $500 a month pooled for Team subscribers. Together the three moves lower the cost of running a multi-model agent, which is the use case Anthropic is selling. (Reuters)

The comparison Anthropic reaches for is OpenAI's GPT-6 Luna, the small OpenAI model it benchmarks against directly. That framing puts Haiku 5.5 against OpenAI's small tier, where the two labs now compete most directly. It also signals where Anthropic expects the volume to come from: cheap calls made many times per task, not a single frontier request.

The launch is the third Claude 5.5 model in about a month. Sonnet 5.5 and Opus 5.5 came first, and Reuters frames the sequence as Anthropic filling out a lineup ahead of a planned IPO. A complete family matters to a company heading toward public markets: it gives enterprise buyers a tiered menu, from the frontier down to the cheap subagent layer, and it gives investors a story in which every price point is covered. OpenAI's GPT-6 Luna occupies the same small tier, which is why Anthropic benchmarks Haiku 5.5 against it at all. Both companies are now selling a cheap model into the same developer market.

Haiku 5.5 is available on the Claude API and through Amazon Web Services (Bedrock), Google Cloud, and Microsoft Foundry on Azure, alongside the same-day pricing changes listed in Anthropic's release notes. (Anthropic release notes)

Caveats

Almost every performance figure in this story comes from Anthropic. They are self-reported vendor numbers, not results from an independent evaluation, and they were measured against competitors on tests Anthropic chose. The claim that Haiku 5.5 costs about 75% less to run on average than Haiku 4.5 is Anthropic's, as is the 90% and 50% split and the tokenizer explanation behind it, though Reuters reported the same roughly 75% headline figure. The cybersecurity safeguard description, including the comparison to Haiku 4.5 and Sonnet 5.5, is Anthropic's account of its own policy. The planned IPO is planned, not completed, and Anthropic remains private. What is missing is independent benchmark data on Haiku 5.5's real-world cost in production, which is the number that would matter most to a team deciding whether to move work onto it.

What to watch

Three things will test the launch. First, whether independent benchmark indexes place Haiku 5.5 near the numbers Anthropic published or below them. Second, whether developers report that the roughly 75% cost saving holds on their own workloads, given the new tokenizer. Third, whether the once-a-month 5.5 cadence continues toward the IPO, and whether Anthropic's filing eventually puts any of these claims in front of a regulated audience. Until then, the clearest fact is the price: at $0.10 for input and $0.50 for output per million tokens on most requests, Haiku 5.5 is the cheapest model Anthropic sells, by its own account.

Cover image: Department for Science, Innovation and Technology, Wikimedia Commons, CC BY 2.0.