There’s a specific frustration that hits when you’re building something that needs to run thousands of times a day. The model that’s good enough for the task is also expensive enough to make the unit economics hurt. You start making compromises you didn’t want to make. You throttle call frequency. You batch things that shouldn’t be batched. You start wondering whether the thing you’re building can actually be a business.
Haiku exists to solve that problem. And Claude Haiku 5.5, announced October 7, 2026, is the most capable version of it Anthropic has ever shipped.
Here’s the full breakdown: what it is, what it costs, what the benchmarks say, and where it actually fits versus its bigger siblings.
What Is Claude Haiku 5.5?
Claude Haiku 5.5 is Anthropic’s fastest, cheapest model in the Claude 5.5 family, which also includes Claude Sonnet 5.5 and Claude Opus 5.5. Per Anthropic’s announcement, it’s designed specifically for high-volume, cost-sensitive work: summarization, subagent tasks, database queries, classification, browser use, and anything else you need to run at a scale that would make a larger model’s price tag unsustainable.
The model ID is claude-haiku-5-5. It’s available now on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure.
Two things separate Haiku 5.5 from Haiku 4.5 beyond the obvious capability jump: the price is dramatically lower, and it’s the first Haiku-class model to ship with an adjustable effort setting. More on both shortly.
Claude Haiku 5.5 Pricing: How Much Does It Cost?
This is probably what you actually came here for, so let’s get to it.
Haiku 5.5 has a two-tier pricing structure based on prompt length. Anthropic designed it this way based on the fact that around 90% of requests to the previous Haiku model fell under 100,000 tokens. The table below is directly from Anthropic’s announcement:
| Token type | Haiku 5.5 (up to 100K tokens) | Haiku 5.5 (over 100K tokens) | Haiku 4.5 |
|---|---|---|---|
| Input tokens | $0.10/M | $0.50/M | $1.00/M |
| Output tokens | $0.50/M | $2.50/M | $5.00/M |
| Cache reads | $0.01/M | $0.05/M | $0.10/M |
| Cache writes (5-minute) | $0.125/M | $0.625/M | $1.25/M |
| Cache writes (1-hour) | $0.20/M | $1.00/M | Not published in this comparison |
For prompts up to 100,000 tokens, Haiku 5.5’s listed input, output, and cache-read rates are 90% lower than Haiku 4.5’s. For requests over 100K tokens, it’s 50% lower. Anthropic estimates an average overall cost reduction of approximately 75%, accounting for the tokenizer change described below.
The 90% reduction on cache reads ($0.01 per million versus $0.10) is the number that matters most for agentic and pipeline workloads, where cached context makes up a large share of total input tokens. At that price, context-heavy repeated calls get dramatically cheaper.
One material technical note from Anthropic’s migration documentation: the same input text produces approximately 30% more tokens with Haiku 5.5 than with Haiku 4.5, although the exact increase varies by content type. This is due to an updated tokenizer similar to Sonnet 5.5’s and Opus 5.5’s. The lower pricing more than offsets this for most workloads, but it’s worth factoring into any cost model before migrating.
Pricing per Anthropic’s official October 7, 2026 announcement. Always verify current rates at the Claude Platform before building production cost models.
Benchmarks: What Claude Haiku 5.5 Can Actually Do
All benchmark figures below are from Anthropic’s official announcement. None have been independently reproduced by external researchers as of the publication date.
| Benchmark | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 (reference) |
|---|---|---|---|---|
| GDPval-AA v2.1 (knowledge work) | 1620 | 735 | 1437 | 1844 |
| AA-Briefcase v1.1 (knowledge work) | 1578 | 614 | 1336 | 1811 |
| OSWorld 2.1, offline subset (computer use) | 72.4% | 15.7% | 48.9% | 83.9% |
| Humanity’s Last Exam, no tools | 45.9% | 10.2% | N/A | 56.9% |
| Humanity’s Last Exam, with tools | 57.4% | 18.7% | N/A | 64.5% |
| Terminal-Bench 4.0 (agentic coding) | 39.2% | 0.0% | 16.4% | 70.6% |
| FrontierCode 1.1 Main (coding) | 46.4% | N/A | 42.4% | 52.1% |
| Chartography (visual reasoning) | 46.4% | 6.4% | 29.1% | 61.6% |
A few things worth reading carefully in these numbers.
The jump from Haiku 4.5 to Haiku 5.5 is not incremental. On GDPval-AA (knowledge work), Haiku 5.5 more than doubles Haiku 4.5’s score. On OSWorld (computer use), it goes from 15.7% to 72.4%. On Terminal-Bench 4.0 under the reported evaluation setup, Haiku 4.5 scored 0.0% and Haiku 5.5 scores 39.2%. These are model-generation-level jumps, not tuning adjustments.
In Anthropic’s published comparisons, Haiku 5.5 scores higher than GPT-6 Luna on the listed GDPval-AA, AA-Briefcase, OSWorld, Terminal-Bench, and Chartography evaluations. These results reflect the specific benchmarks Anthropic published rather than every task in those capability areas.
The Sonnet 5.5 column is there for reference. Haiku 5.5 is not a replacement for Sonnet 5.5. On Terminal-Bench 4.0 especially, the gap is significant: 39.2% versus 70.6%. Complex agentic coding tasks still belong with Sonnet or Opus. What Haiku 5.5 does is unlock a class of work that was previously cost-prohibitive, not replace the models above it.
The Effort Setting: Haiku 5.5’s New Dial
Haiku 5.5 is the first Haiku-class model to include an adjustable effort parameter. Per Anthropic’s announcement, this means users can optimize for either cost or intelligence depending on the task.
This follows the same structure as Sonnet 5.5 and Opus 5.5, which already support adjustable effort. The practical implication: you can run Haiku 5.5 at low effort for mechanical high-volume work, ramp to medium for tasks that need more reasoning, or push higher for the complex edge cases in an otherwise cheap pipeline.
Anthropic’s published charts show how this plays out on OSWorld, GDPval-AA, and Humanity’s Last Exam at each effort level (Low, Med, High, Xhigh, Max), plotted against cost per attempt. The short version: at low effort, Haiku 5.5 is extremely cheap and still meaningfully capable. At higher effort settings, it extends meaningfully into use cases where you’d previously default to Sonnet.
Technical Specifications
| Spec | Claude Haiku 5.5 |
|---|---|
| Model ID | claude-haiku-5-5 |
| Context window | 1 million tokens |
| Maximum output | 128K tokens |
| Vision / multimodal | Yes |
| Computer use | Yes |
| Effort setting | Yes (adjustable) |
| Tool use | Yes |
| Release date | October 7, 2026 |
The 1 million token context window and 128K output limit match Sonnet 5.5 and Opus 5.5. That’s a meaningful upgrade from what previous Haiku models offered and makes Haiku 5.5 viable for workloads that require long context without requiring a larger, more expensive model.
What Claude Haiku 5.5 Is Actually Built For
Anthropic is explicit about the target workloads, and it’s worth being equally explicit here rather than letting “small model” imply limited usefulness.
Subagents. This is where Haiku 5.5 fits most cleanly in multi-model architectures. Rogo’s team described the workflow directly: while a larger model builds the deck, a Haiku 5.5 subagent goes into the 10-K and pulls the specific data the deck needs. Accurate enough to trust on that task. Fast and cheap enough to run constantly. Cognition confirmed a similar pattern: in Devin Fusion, Haiku 5.5 as the sidekick alongside Opus 5.5 as the lead model held a top-tier FrontierCode score of 66.2 while cutting cost and latency.
Summarization and compaction. AlphaSense runs approximately 8 million calls per week on their “Ask in Document” feature, answering specific questions on top of one or a few documents. Per their team’s report, Haiku 5.5 showed a statistically significant improvement over Haiku 4.5 at that volume (0.84 versus 0.76 on their internal evaluation). At 8 million calls per week, the 75% average cost reduction is not a minor line item.
Browser use and computer use. Haiku 5.5 scores 72.4% on OSWorld 2.1’s offline subset, which measures how well an agent can operate a real computer to finish multi-step tasks. That’s an extremely strong result for a model at this price tier. Alongside this launch, Anthropic updated the Claude Python and TypeScript SDKs to add beta support for computer use and browser use, and they explicitly call out Haiku 5.5 as “especially well-suited to these tasks, given its combination of speed, capability, and price.”
Live customer support. Anthropic describes Haiku 5.5 as their fastest model to date at each model’s standard speed. Asana’s engineering team confirmed a 30% reduction in latency for task completions and up to 2.5x faster inference per agent turn. For real-time interactions where response speed is directly visible to users, that matters.
High-volume CRM and classification. HubSpot tested Haiku 5.5 on simulated portals for CRM tasks including deal reporting. It scored 92.8% averaged over three runs on their evaluation suite, the highest score they’d seen for that benchmark. On one specific task (identifying stale but ambiguous records), it was the fastest to complete and had the highest hit rate and lowest false positive rate of all models tested.
What Claude Haiku 5.5 Is Not For
The same announcement that introduces Haiku 5.5 includes a chart that makes this clear without ambiguity. On Terminal-Bench 4.0, which measures complex multi-step professional tasks in a command-line interface, Haiku 5.5 scores 39.2%. Sonnet 5.5 at low effort still outperforms it. Sonnet 5.5 at higher effort settings widens that gap considerably.
Anthropic’s phrasing is worth quoting directly: “Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding tasks like those measured by Terminal-Bench 4.0. By contrast, Haiku 5.5 is best suited to more narrowly scoped tasks that might otherwise have been cost-prohibitive with previous versions of Claude.”
So: Haiku 5.5 for scoped, high-volume, speed-sensitive tasks. Sonnet 5.5 and Opus 5.5 for complex, long-horizon agentic work. The two tiers are designed to be used together, not to replace each other.
Claude Haiku 5.5 vs Claude Sonnet 5.5 vs Claude Opus 5.5
| Factor | Haiku 5.5 | Sonnet 5.5 | Opus 5.5 |
|---|---|---|---|
| Input (up to 100K) | $0.10/M | $2.00/M | $4.00/M |
| Output (up to 100K) | $0.50/M | $10.00/M | $20.00/M |
| Cache reads | $0.01/M | $0.10/M | $0.20/M |
| Terminal-Bench 4.0 | 39.2% | 70.6% | 66.4% |
| GDPval-AA | 1620 | 1844 | 1846 |
| OSWorld 2.1 | 72.4% | 83.9% | 81.8% |
| Speed | Fastest | Fast | Standard |
| Best for | Subagents, summaries, browser use, classification | Complex coding, knowledge work | Hardest agentic tasks, long-horizon work |
Sonnet 5.5 cache read pricing as of October 7, 2026 ($0.10/M, reduced 50% from launch). Opus 5.5 prices per Anthropic’s published rates. All benchmark figures from Anthropic’s official announcements.
Claude Haiku 5.5 vs Haiku 4.5
The migration case here is unusually clear. Haiku 5.5 is 90% cheaper on the vast majority of requests. It outperforms Haiku 4.5 by multiples on every published benchmark. It supports adjustable effort, which Haiku 4.5 did not. And it adds computer use capability, where Haiku 4.5 scored 0.0% on Terminal-Bench.
The one thing to account for in the migration is the tokenizer change. Haiku 5.5 uses slightly more tokens per task due to its updated tokenizer. Anthropic has published a migration guide at platform.claude.com/docs/en/models/haiku-5-5/migration-guide with specifics.
For teams running high-volume Haiku 4.5 workloads, the economics of staying on Haiku 4.5 are hard to justify.
The Sonnet 5.5 Cache Read Price Cut (Announced Alongside Haiku 5.5)
Worth knowing even if Haiku 5.5 is your primary interest: Anthropic simultaneously cut Sonnet 5.5’s cache read pricing by 50%, from $0.20 to $0.10 per million tokens. Because cache reads make up a large share of total token consumption in agentic workloads, Anthropic says this reduces the cost of Sonnet 5.5 on most agentic tasks by around 20%.
This changes the cost-performance math for teams deciding between Haiku 5.5 and Sonnet 5.5 for mid-complexity tasks. Sonnet 5.5 just got meaningfully cheaper for any workload with substantial cached context.
Monthly API Credits for Max and Team Subscribers
Also announced October 7: Anthropic is rolling out a new monthly API credit to Max and Team subscribers for use on the Claude Platform.
Per Anthropic’s announcement: Max 5x users get $100 per month, Max 20x users get $200 per month, and Team subscribers get credits pooled across their users, up to $500 depending on the number and type of seats. These credits apply to the Claude Platform API and are designed for building tools, apps, and agents. Free and Pro plans are not eligible. For full eligibility details, see Anthropic’s Help Center article.
This is directly relevant to Haiku 5.5’s target audience: developers and teams building agentic applications who want to experiment before committing to production-scale API spend.
Safety
Per Anthropic’s announcement, Haiku 5.5 shows major improvements across almost all alignment evaluations relative to Haiku 4.5, including fewer instances of misaligned behavior and lower willingness to cooperate with misuse.
On safeguards: Haiku 5.5’s cybersecurity guardrails are more restrictive than Haiku 4.5’s but somewhat less restrictive than Sonnet 5.5’s. They permit a wider range of defensive security tasks than Sonnet 5.5, but still block penetration testing and techniques more likely to be used by attackers. Biology safeguards match Sonnet 5, Sonnet 5.5, and Opus 5.
Organizations that need broader access to biology or cybersecurity capabilities can apply to Anthropic’s Life Sciences Verification Program or Cyber Verification Program.
Full evaluation details are in the Haiku 5.5 System Card.
Availability and How to Access Claude Haiku 5.5
Haiku 5.5 is available now on all platforms. On the Claude Platform, use model ID claude-haiku-5-5. It’s also available on Amazon Web Services (Bedrock), Google Cloud (Vertex AI), and Microsoft Azure (AI Foundry).
API access: Available on the Claude Platform via claude-haiku-5-5. Requires a Claude Platform account. Consumer subscription availability varies by plan; check claude.ai for current plan details.
Migration guide: platform.claude.com/docs/en/models/haiku-5-5/migration-guide
Frequently Asked Questions
What is Claude Haiku 5.5?
Claude Haiku 5.5 is Anthropic’s latest small model, announced October 7, 2026. Per Anthropic’s description, it’s their fastest, cheapest, and most capable small model to date, designed for high-volume tasks like summarization, browser use, subagent work, and classification. Model ID: claude-haiku-5-5.
How much does Claude Haiku 5.5 cost?
For prompts up to 100,000 tokens: $0.10 per million input tokens, $0.50 per million output tokens, $0.01 per million cache reads, and $0.125 per million for 5-minute cache writes ($0.20 per million for 1-hour cache writes). For prompts over 100,000 tokens: $0.50 input, $2.50 output, $0.05 cache reads per million tokens. All rates are substantially lower than Haiku 4.5. Verify current rates at claude.com/pricing.
When was Claude Haiku 5.5 released?
October 7, 2026.
What is Claude Haiku 5.5’s context window?
1 million tokens input context, 128K token maximum output, per Anthropic’s published specifications.
How does Claude Haiku 5.5 compare to Haiku 4.5?
Dramatically better on every published benchmark. On GDPval-AA (knowledge work), Haiku 5.5 more than doubles Haiku 4.5’s score. On OSWorld (computer use), it jumps from 15.7% to 72.4%. On Terminal-Bench 4.0 under the reported evaluation setup, Haiku 4.5 scored 0.0% and Haiku 5.5 scores 39.2%. Listed per-token rates are 90% lower for prompts under 100K tokens, though the updated tokenizer produces approximately 30% more tokens for the same input, so the effective cost savings vary by workload. Anthropic’s overall estimate is approximately 75% lower average running cost.
Is Claude Haiku 5.5 good for coding?
For subagent coding tasks, document lookups, and quick targeted operations within a larger agentic system: yes, strong per benchmark results. For complex, multi-step long-horizon software engineering: Sonnet 5.5 and Opus 5.5 remain the better choices, per Anthropic’s own published guidance and the Terminal-Bench gap (39.2% vs 70.6% for Sonnet 5.5).
Does Claude Haiku 5.5 support computer use?
Yes. It scores 72.4% on OSWorld 2.1’s offline subset. Alongside this launch, Anthropic updated the Claude Python and TypeScript SDKs to add beta support for computer use and browser use, calling out Haiku 5.5 as especially well-suited to those tasks.
Is Claude Haiku 5.5 available on AWS, Google Cloud, and Azure?
Yes. Available now on Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure AI Foundry.
What is the effort setting on Haiku 5.5?
Haiku 5.5 is the first Haiku-class model to include an adjustable effort parameter, letting users optimize for cost or intelligence depending on the task. It supports Low, Med, High, Xhigh, and Max effort levels, consistent with Sonnet 5.5 and Opus 5.5.
Is Claude Haiku 5.5 free?
API access is pay-per-token with no free tier. Consumer subscription access varies by plan; check claude.ai for current details. Max and Team subscribers receive monthly API credits ($100 to $500 depending on plan and seats) usable on any Claude model including Haiku 5.5. Free and Pro plans are not eligible for the monthly API credits.
What is the best use case for Claude Haiku 5.5?
Per Anthropic’s guidance and customer testing: high-volume summarization and compaction, subagent work within multi-model pipelines, browser and computer use automation, live customer support where latency matters, CRM queries, document analysis at scale, and classification tasks. Anything you need to run thousands of times a day at a price that actually works.
How does Claude Haiku 5.5 compare to Claude Sonnet 5.5?
Haiku 5.5 is approximately 20x cheaper on input tokens and 20x cheaper on output tokens compared to Sonnet 5.5. Sonnet 5.5 significantly outperforms Haiku 5.5 on complex agentic coding (Terminal-Bench 4.0: 70.6% vs 39.2%). For scoped, high-volume tasks, Haiku 5.5 is the right choice. For complex reasoning and long-horizon tasks, Sonnet 5.5 is the right choice. The two are designed to pair together.
Final Thoughts
Haiku 5.5 isn’t a small model pretending to be bigger than it is. It’s a genuinely strong model optimized for a specific, commercially important set of tasks, priced aggressively enough that the economics of running it at scale actually make sense.
The benchmark improvements from Haiku 4.5 are dramatic across the board. The price reduction is real and large. And the use cases, subagents, browser automation, summarization pipelines, customer support, live classification, are exactly the workloads that either previously required expensive models or lived in the uncomfortable middle ground of “Haiku 4.5 is almost good enough.”
Almost is over. Haiku 5.5 is good enough for a lot of things that couldn’t justify the cost before. That’s the actual story here.
Start building at platform.claude.com with model ID claude-haiku-5-5.
All benchmark figures, pricing, customer quotes, and technical specifications in this article are sourced from Anthropic’s official Claude Haiku 5.5 announcement at anthropic.com/claude-haiku-5-5, published October 7, 2026. Benchmark scores are from Anthropic’s own evaluations and have not been independently reproduced by external researchers as of publication. Pricing is API list pricing and may differ from effective rates on subscription plans. Always verify current pricing at claude.com/pricing before building production cost models.
Curated byย Lorphic
Digital intelligence. Clarity. Truth.
