Claude Haiku 5.5: The Enterprise Subagent Model Is Here
Claude Haiku 5.5: The Enterprise Subagent Model Is Here
On October 7, 2026, Anthropic released Claude Haiku 5.5, the fastest and cheapest model in the Claude 5.5 family. The headline: it costs roughly 75% less than Haiku 4.5 for most enterprise tasks, and for the first time the Haiku tier is a credible choice for computer use and agentic coding loops, not just routing and classification. For teams building multi-agent systems, the release redraws the cost map for subagent infrastructure.
What Changed
Haiku 5.5 completes the Claude 5.5 family launched with Opus 5.5 in September 2026. Three things are genuinely new compared to Haiku 4.5:
Adaptive thinking. Haiku 5.5 is the first Haiku model with an adjustable effort parameter, running from low through max. This lets teams tune cost against capability per task type rather than accepting a single fixed setting across a whole pipeline. The default on the Claude API and in Claude Code is medium.
Computer use and terminal agents. The numbers below tell the story. At max effort, Haiku 5.5 scored 72.4% on Anthropic’s OSWorld 2.1 offline partial benchmark, compared to 15.7% for Haiku 4.5. On Terminal-Bench 4.0, it scored 39.2% versus a literal 0.0% for its predecessor. These gains close the gap that made the previous Haiku a non-starter for agent loops.
Two-tier pricing. The $0.10/$0.50 rate applies to prompts under 100,000 tokens. Roughly 90% of Haiku 4.5 requests fell under that threshold, so the 75% average cost reduction holds for most existing workloads. Prompts over 100,000 tokens move to $0.50/$2.50. Note: Haiku 5.5 uses Anthropic’s newer tokenizer, which counts about 30% more tokens for the same text, so near-threshold workloads should be recounted before budgeting.
Benchmark Snapshot
The table below compares Haiku 5.5 (at max effort) against Haiku 4.5, GPT-6 Luna, and Sonnet 5.5 across the benchmarks Anthropic published at launch. All figures are self-reported by Anthropic.
| Benchmark | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|
| GDPval-AA v2.1 (Elo) | 1,620 | 735 | 1,437 | 1,840 |
| AA-Briefcase v1.1 (Elo) | 1,578 | 614 | 1,336 | 1,824 |
| OSWorld 2.1 (partial) | 72.4% | 15.7% | 48.9% | 83.9% |
| Terminal-Bench 4.0 | 39.2% | 0.0% | 16.4% | 70.6% |
| FrontierCode 1.1 Main | 46.4% | n/a | 42.4% | 52.1% (xhigh) |
| Humanity’s Last Exam (no tools) | 45.9% | 10.2% | n/a | 56.9% |
| Chartography (with tools) | 86.2% | n/a | n/a | 90.2% |
Artificial Analysis’s independent evaluation confirms comparable scores, placing Haiku 5.5 at an Intelligence Index of 43, ahead of GPT-6 Luna (38) and Gemini 3.8 Flash (41) at max effort. It trails Sonnet 5.5 by around 13 points on that index.
The key read for enterprise builders: Haiku 5.5 beats GPT-6 Luna on OSWorld partial (72.4% vs 48.9%) and Terminal-Bench (39.2% vs 16.4%), the two benchmarks that matter most for computer use and agentic coding subagent tasks.
The Enterprise Use Case: Subagent Orchestration
The canonical deployment pattern for Haiku 5.5 is the two-model pipeline. Opus 5.5 or Sonnet 5.5 handles planning, judgment, and complex reasoning. Haiku 5.5 handles execution of well-scoped subtasks: routing requests, summarizing chunks, reviewing file diffs, running queries, and operating browsers or desktops at high volume.
Cognition’s Devin Fusion demonstrated this pattern publicly at launch: with Haiku 5.5 as a sidekick, Fusion holds a FrontierCode score of 66.2 in the Devin CLI while reducing cost and latency versus a single-model setup.
Customer benchmarks from the launch announcement reinforce this:
- Asana reported over 30% latency reduction and up to 2.5x faster inference per agent turn versus their prior model in an AI Teammates evaluation covering bug triage, project setup, and portfolio risk analysis.
- HubSpot measured 92.8% on their CRM suite (averaged over three runs), the best score they had seen on the suite, including tasks for identifying stale deal records and CRM audits requiring Haiku to distinguish ambiguous cases.
- Box saw Haiku 5.5 score 11 points higher than Haiku 4.5 at about half the latency on enterprise content analytical tasks, including cost reports and financial summaries.
- AlphaSense recorded 0.84 versus 0.76 accuracy on 400 document-specific queries using their Ask in Document feature, which runs approximately 8 million calls per week in production.
These are customer-reported figures from Anthropic’s launch page, not independently verified. They nonetheless illustrate the consistent pattern: Haiku 5.5 improves over Haiku 4.5 on real production workloads where speed and accuracy both matter.
Full Pricing Reference
The table below shows how Haiku 5.5 fits into the current Claude 5.5 lineup as of October 2026. Prices are per million tokens on the Claude API.
| Model | Input (short) | Output (short) | Input (long) | Output (long) | Cache read |
|---|---|---|---|---|---|
| Claude Haiku 5.5 | $0.10 | $0.50 | $0.50 | $2.50 | $0.01 |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $2.00 | $10.00 | $0.10 |
| Claude Opus 5.5 | $4.00 | $20.00 | $4.00 | $20.00 | $0.40 |
| Claude Fable 5.1 | $10.00 | $50.00 | $10.00 | $50.00 | $1.00 |
“Short” applies to prompts under 100,000 tokens; “long” applies above. Sonnet 5.5 has a single tier. Batch API pricing adds a 50% discount across all models. Prompt caching saves up to 90% on cache reads.
Note the Sonnet 5.5 cache read change: Anthropic cut it from $0.20 to $0.10 at the same time as the Haiku 5.5 launch, reducing Sonnet 5.5’s cost on most agentic tasks by approximately 20%. For teams running high-context agentic loops on Sonnet 5.5, this is a meaningful reduction that happens automatically with no code changes.
Availability
Claude Haiku 5.5 is available immediately on:
- Claude API (model string:
claude-haiku-5-5) - Amazon Bedrock (US, EU, AU, JP, Global inference profiles, including AWS GovCloud)
- Google Cloud Vertex AI
- Microsoft Foundry
- Claude Platform on AWS
- Claude.ai on all plans (Free, Pro, Max, Team, Enterprise) on web, iOS, and Android
- Claude Code (default
mediumeffort)
Anthropic also updated its Python and TypeScript SDKs to add support for computer use and browser use in beta with this release. Haiku 5.5’s combination of speed, computer use capability, and price makes it the natural default for browser and desktop automation steps in agent pipelines where Sonnet 5.5 would otherwise be underused.
What This Means for Enterprise AI Strategy
For enterprise teams building AI-native agent infrastructure, Haiku 5.5 changes a specific calculation: the cost floor for computer use and terminal tasks. Previous to this release, Haiku was too weak for these tasks, meaning teams either used Sonnet (expensive at scale) or skipped the Haiku tier entirely.
With Haiku 5.5 scoring 72% on OSWorld and 39% on Terminal-Bench, there is now a credible cheap-layer option for browser automation, file operations, and basic code review subtasks in a multi-agent pipeline. The right pattern is selective: use Haiku 5.5 where the task is well-scoped and repetitive, keep Sonnet 5.5 or Opus 5.5 in the lead role for planning and judgment calls.
The practical limit remains real. Sonnet 5.5 still scores 83.9% on OSWorld versus Haiku’s 72.4%, and 70.6% on Terminal-Bench versus Haiku’s 39.2%. For tasks where computer use errors are costly, Sonnet remains the safer choice. The Haiku tier earns its place in high-volume, low-stakes automation and as an execution layer in multi-agent orchestration, not as a general-purpose replacement for the mid-tier.
For teams evaluating the shifting economics of enterprise AI agent infrastructure, Haiku 5.5 plus the Sonnet 5.5 cache read cut represents roughly a 20 to 40% total cost reduction on well-optimized agentic workloads, depending on the mix of models and prompt sizes. That is a meaningful reduction for teams already running agents at scale.
Claude Haiku 5.5 is available now. The model string is claude-haiku-5-5 across all platforms. Anthropic has committed to keeping the model available for at least one year, with no planned retirement before October 7, 2027.