Claude Haiku 5.5: Anthropic's Cheapest, Fastest Small Model With Adjustable Effort, 1M Context, and Browser Use

Executive Summary

On October 7, 2026, Anthropic released Claude Haiku 5.5, describing it as the "cheapest, fastest, and most capable small model" it has ever built. Haiku 5.5 is designed for high-volume, cost-sensitive workloads — summaries, compactions, database queries, classification, and subagent tasks — and represents a fundamental shift in the economics of small-model inference.

At $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens, Haiku 5.5 costs an average of 75% less than Haiku 4.5 and matches OpenAI's GPT-6 Luna on list pricing. The model supports a 1M-token context window, is the first Haiku with adjustable effort settings, and delivers significant performance gains: 1620 on GDPval-AA v2.1 (vs. 1162 for Haiku 4.5), 72.4% on OSWorld 2.1 (vs. 15.7%), and 46.4% on FrontierCode 1.1 (vs. 42.4% for GPT-6 Luna).

The release is accompanied by three additional value improvements: Sonnet 5.5 cache reads halved to $0.10/M (reducing Sonnet 5.5's cost on most agentic work by ~20%), new monthly API credits for Max and Team subscribers ($100–$500/month), and beta support for computer use and browser use in the Claude Python and TypeScript SDKs — with Haiku 5.5 specifically recommended for these latency-sensitive tasks.

What makes Haiku 5.5 notable isn't just the price — it's the strategic repositioning of the small-model tier. By adding 1M context, adjustable effort, and browser use to a model that costs $0.10/M input, Anthropic is making it economically viable to deploy agents at scale for tasks that were previously too expensive to run on Claude.

---

1. What Is Haiku 5.5?

1.1 The Small-Model Tier, Reimagined

Haiku 5.5 is Anthropic's latest entry in the Haiku class — models designed for speed, volume, and cost efficiency. But it represents a qualitative leap over previous Haiku models:

Feature	Haiku 3.5	Haiku 4.5	Haiku 5.5
Context window	200K	200K	1M
Adjustable effort	❌	❌	✅ (first Haiku)
Input pricing	$0.25/M	$1.00/M	$0.10/M
Output pricing	$1.25/M	$5.00/M	$0.50/M
Browser use	❌	❌	✅ (beta)
Computer use	❌	❌	✅ (beta)
Claude Code	❌	❌	✅

The 1M context window is the most dramatic architectural change — a 5× increase over the previous Haiku generation. Combined with the $0.10/M input price, this makes Haiku 5.5 viable for long-document processing at scale, a use case that was previously cost-prohibitive for the Haiku tier.

1.2 The Adjustable Effort Breakthrough

Haiku 5.5 is the first Haiku-class model with adjustable effort settings, bringing it into parity with Sonnet and Opus on this dimension. Users can now tune the cost-intelligence trade-off per request:

"It is also the first Haiku with effort controls, so teams can tune cost against intelligence for each task."

This is significant because it means Haiku 5.5 can now be used in multi-tier agent architectures where a smarter model (Opus 5.5 or Sonnet 5.5) plans the work and hands off subtasks to Haiku with effort settings matched to task difficulty:

1.3 Designed for Volume, Not Complexity

Anthropic is clear about Haiku 5.5's positioning:

"Haiku 5.5 is best suited to more narrowly scoped tasks that might otherwise have been cost-prohibitive with previous versions of Claude — like compaction, summarization, or subagent work."

The model is not intended to replace Sonnet 5.5 or Opus 5.5 for complex agentic coding tasks. On Terminal-Bench 4.0, Haiku 5.5 scores 39.2% vs. Sonnet 5.5's 70.6% — a significant gap that confirms the tiered architecture. But for the tasks it's designed for, it's dramatically more capable than Haiku 4.5.

---

2. Benchmark Performance: The Small-Model Leap

2.1 Anthropic's Official Comparison

Anthropic published a comprehensive comparison table pitting Haiku 5.5 against Haiku 4.5, GPT-6 Luna, and Sonnet 5.5:

Benchmark	Haiku 5.5	Haiku 4.5	GPT-6 Luna	Sonnet 5.5	Direction
GDPval-AA v2.1 (knowledge work)	1620	1162	1351	1840	Higher = better
AA-Briefcase v1.1 (office work)	1578	1413	1336	1824	Higher = better
OSWorld 2.1 (computer use, offline)	72.4%	15.7%	48.9%	83.9%	Higher = better
HLE (expert reasoning, no tools)	45.9%	10.2%	—	56.9%	Higher = better
HLE (expert reasoning, with tools)	57.4%	18.7%	—	64.5%	Higher = better
Terminal-Bench 4.0 (agentic coding)	39.2%	0.0%	16.4%	70.6%	Higher = better
FrontierCode 1.1 (coding)	46.4%	—	42.4%	52.1%	Higher = better
Chartography (visual reasoning)	46.4%	6.4%	29.1%	61.6%	Higher = better

2.2 Where Haiku 5.5 Leads

The most dramatic improvements are in computer use and visual reasoning:

• OSWorld 2.1: A 361% improvement from 15.7% to 72.4%, surpassing GPT-6 Luna (48.9%) by 23.5 points. This is the single largest gain in the comparison.
• HLE (no tools): A 350% improvement from 10.2% to 45.9%.
• Chartography: A 625% improvement from 6.4% to 46.4%, surpassing GPT-6 Luna (29.1%) by 17.3 points.
• Terminal-Bench 4.0: A 39.2-point improvement from 0.0% to 39.2%, surpassing GPT-6 Luna (16.4%) by 22.8 points.

2.3 The Effort Setting Impact

Anthropic published accuracy-vs-cost charts for OSWorld 2.1 and Terminal-Bench 4.0 across effort settings, showing that Haiku 5.5's performance scales with effort investment. The charts demonstrate that:

• Low effort: Best for simple, repetitive tasks where speed matters most
• Medium effort: Balanced cost-intelligence for most subagent work
• High effort: Approaches Sonnet 5.5 performance on narrower tasks at a fraction of the cost

Early customer testing by Asana reported over 30% reduction in latency and up to 2.5× faster inference per agent turn compared to their current model, with results consistent across their eval suite for AI Teammates.

---

3. Pricing: The $0.10/M Floor

3.1 Token Pricing

Price per 1M tokens	Haiku 5.5 (≤100K)	Haiku 5.5 (>100K)	Haiku 4.5	Sonnet 5.5	GPT-6 Luna
Input	$0.10	$0.50	$1.00	$2.00	$0.10
Cached input	$0.01	$0.05	$0.10	$0.10	$0.20
Output	$0.50	$2.50	$5.00	$10.00	$0.50

Haiku 5.5 matches GPT-6 Luna on list pricing ($0.10/$0.50) but offers a 1M context window (vs. 1.05M for Luna) and adjustable effort (which Luna lacks in the Haiku tier). The cached input price of $0.01/M is the lowest of any Claude model — a 90% discount from standard input.

3.2 The 100K Token Threshold

A critical detail: the $0.10/$0.50 pricing applies only to prompts up to 100K tokens. For prompts exceeding 100K tokens, pricing jumps to $0.50/$2.50 — a 5× increase. Anthropic notes that prompts up to 100K tokens represent ~90% of requests to the previous Haiku model, suggesting most workloads will stay in the lower tier.

3.3 Real-World Cost Impact

Anthropic estimates that workloads cost approximately 75% less to run on Haiku 5.5 compared to Haiku 4.5, accounting for changes in request sizes and token consumption patterns. This is a dramatic improvement that could transform the economics of high-volume agent deployments.

---

4. Key Features

4.1 1M Context Window

Haiku 5.5's 1M-token context window is a 5× increase over the previous Haiku generation and matches the context length of Sonnet 5.5 and Opus 5.5. This enables:

• Long-document processing at scale (10-Ks, legal contracts, technical specifications)
• Extended conversation history for agent workloads
• Full-codebase awareness for subagent coding tasks

4.2 Adjustable Effort

The first Haiku with effort controls, Haiku 5.5 allows developers to:

• Optimize for cost on simple, repetitive tasks (low effort)
• Balance cost and intelligence for standard subagent work (medium effort)
• Maximize capability on harder tasks (high effort)

This brings Haiku 5.5 into parity with the rest of the Claude 5.5 family on this dimension and enables fine-grained cost control in multi-agent architectures.

4.3 Browser Use and Computer Use (Beta)

Anthropic is rolling out computer use and browser use support in the Claude Python and TypeScript SDKs, with Haiku 5.5 specifically recommended for these tasks:

"Haiku 5.5 is especially well-suited to these tasks, given its combination of speed, capability, and price."

The browser use SDK enables agents to:

• Fill forms and navigate web applications
• Extract data from dynamic web pages
• Perform repetitive browser-based workflows
• Integrate with real-time customer support systems

4.4 Claude Code Integration

Haiku 5.5 is available in Claude Code, Anthropic's coding agent, making it accessible for developers who want to use the model for simple coding tasks, code reviews, and automated refactoring.

---

5. Safety and Alignment

5.1 Alignment Improvements

Haiku 5.5 shows major improvements across almost all alignment evaluations relative to Haiku 4.5:

"We found far fewer instances of misaligned behavior, and a lower willingness to cooperate with misuse."

The full evaluation methodology and results are documented in the Haiku 5.5 System Card.

5.2 Cybersecurity Safeguards

Haiku 5.5's cybersecurity safeguards are more restrictive than Haiku 4.5 but somewhat less restrictive than Sonnet 5.5:

• Permits: A wider range of defensive cybersecurity tasks
• Blocks: Penetration testing and techniques more likely to be used by attackers
• Verification programs: Organizations working on wider-ranging cyber activities can apply to the Cyber Verification Program

5.3 Biology Safeguards

Haiku 5.5's biology safeguards are the same as Sonnet 5, Sonnet 5.5, and Opus 5:

• Allows: Research biology questions
• Restricts: Requests judged as likely to cause harm
• Verification programs: Organizations can apply to the Life Sciences Verification Program

---

6. Deployment and Access

6.1 Availability Channels

Channel	Status	Notes
Claude Platform	✅ Available	claude-haiku-5-5 model slug
Claude.ai	✅ Available	Web, iOS, Android
Amazon Web Services	✅ Available	Through Bedrock
Google Cloud	✅ Available	Through Vertex AI
Microsoft Azure	✅ Available	Through Azure AI
Microsoft Foundry	✅ Available	Enterprise deployment
Claude Code	✅ Available	Coding agent integration
Third-party providers	✅ Available	OpenRouter, etc.

6.2 Migration from Haiku 4.5

Anthropic provides a migration guide for transitioning from Haiku 4.5 to Haiku 5.5, covering:

• API endpoint changes
• Tokenizer differences (which can affect the 100K threshold)
• Effort setting configuration
• Prompt caching optimization

6.3 New API Credits for Subscribers

Starting with this release, Anthropic is rolling out monthly API credits for Claude Max and Team subscribers:

Plan	Monthly Credits	Purpose
Max 5×	$100/month	Experiment with agents and apps
Max 20×	$200/month	Experiment with agents and apps
Team	$500/month (pooled)	Experiment with agents and apps

These credits can be used on any Claude model and are designed to support users building new agents and applications on the Claude Platform.

---

7. Customer Results

7.1 Asana: 30% Latency Reduction

Asana ran Haiku 5.5 through their eval suite for AI Teammates, covering bug triage, project setup, and portfolio search:

"Compared with the model we use today, we saw over a 30% reduction in latency for task completions and up to 2.5× faster inference per agent turn. It's a noticeably snappier experience." — Aaron Vinh, Staff Software Engineer, Asana

7.2 Box: 11-Point Improvement at Half Latency

Box AI tested Haiku 5.5 on analytical work at scale:

"Claude Haiku 5.5 scored 11 points higher than Haiku 4.5 at about half the latency. We'd put it to use on analytical work that runs at scale, from cost reports to financial summaries and weekly recurring reviews."

7.3 HubSpot: 92.8% on CRM Tasks

HubSpot evaluated Haiku 5.5 on simulated CRM portals:

"Claude Haiku 5.5 got the best score we've seen on this suite yet, at 92.8% averaged over three runs. Across all of the models we tested, Haiku 5.5 was fastest to complete the task, and had the highest hit rate and the lowest false positive rate."

7.4 Devin: Top-Tier FrontierCode at Lower Cost

Devin Fusion uses Haiku 5.5 as a sidekick model:

"With Haiku 5.5 as the sidekick, Fusion holds a top-tier FrontierCode score of 66.2 while cutting cost and latency."

---

8. Connection to Prior Research

8.1 The October 2026 Frontier Landscape

Haiku 5.5 arrives in the most active period of the frontier AI race:

• Grok 4 7 Frontier Coding Knowledge Work New Safeguard Stack Longer Rl 2026 10 06 — xAI's Grok 4.7 with longer RL training and new safeguard stack
• Qwen3 8 Omni Flash Native Omnimodal Agentic Media 1m Context 2026 10 05 — Qwen3.8-Omni-Flash with native omnimodal agentic workflows
• Pareto 26 10 Preview Blended Model Pareto Frontier Cost Performance Unbiased 2026 10 02 — Unbiased's Pareto 26.10 Preview with blended-model cost-performance
• Claude Sonnet 5 5 Faster Cheaper Coding Jump Cyber Safeguards Distillation Protection 2026 09 29 — Anthropic's Sonnet 5.5 with 60-point coding jump
• This article — Haiku 5.5 with 1M context, adjustable effort, and $0.10/M pricing

8.2 The Cost-Efficiency Convergence

Haiku 5.5 continues the cost-efficiency narrative established by recent releases:

Model	Input ($/1M)	Output ($/1M)	Context	Positioning
Haiku 5.5	$0.10	$0.50	1M	Small-model, high-volume
GPT-6 Luna	$0.10	$0.50	1.05M	Small-model, high-volume
Pareto 26.10	$0.80	$3.20	1M	Blended ensemble
Qwen3.8-Omni-Flash	$0.15	$0.47	1M	Omnimodal agents
GPT-6.1 Sol	$2.00	$10.00	1.05M	Frontier coding
Sonnet 5.5	$2.00	$10.00	1M	Well-scoped tasks
Grok 4.7	$2.00	$6.00	500K	Coding & knowledge work

Haiku 5.5 and GPT-6 Luna now share the $0.10/M input floor, but Haiku 5.5 differentiates with adjustable effort, browser use, and Claude Code integration.

8.3 The Multi-Tier Agent Architecture

Haiku 5.5 completes Anthropic's three-tier model family:

• Opus 5.5: Planning, orchestration, complex reasoning (highest cost)
• Sonnet 5.5: Well-scoped tasks, agentic coding (mid-tier cost)
• Haiku 5.5: Subagents, high-volume tasks, real-time experiences (lowest cost)

This tiered architecture enables cost-optimized agent systems where each task is handled by the most appropriate model, a pattern that was previewed in the Sonnet 5.5 release (see Claude Sonnet 5 5 Faster Cheaper Coding Jump Cyber Safeguards Distillation Protection 2026 09 29).

---

9. Key Takeaways

1. The $0.10/M floor is now contested. Haiku 5.5 and GPT-6 Luna both offer $0.10/M input pricing, but Haiku 5.5 differentiates with 1M context, adjustable effort, and browser use support.

2. 1M context on a small model changes the game. The 5× context increase over Haiku 4.5 — combined with $0.10/M pricing — makes long-document processing at scale economically viable for the first time in the Haiku tier.

3. Adjustable effort on Haiku enables fine-grained cost control. Being the first Haiku with effort settings, Haiku 5.5 can now be tuned per-task in multi-agent architectures, optimizing the cost-intelligence trade-off dynamically.

4. The 75% cost reduction is transformative. Anthropic estimates 75% lower average costs vs. Haiku 4.5, which could dramatically change the economics of high-volume agent deployments.

5. Browser use and computer use open new use cases. The beta SDK support for browser and computer use — specifically recommended for Haiku 5.5 — enables automated form filling, data extraction, and real-time customer support at scale.

6. Customer results confirm the speed advantage. Asana's 30% latency reduction and 2.5× faster inference, Box's 11-point improvement at half latency, and HubSpot's 92.8% CRM score validate the performance claims.

7. The Sonnet 5.5 cache read cut is a bonus. Halving Sonnet 5.5's cache reads to $0.10/M reduces the cost of most agentic work by ~20%, benefiting users of the mid-tier model.

8. The monthly API credits lower the barrier to entry. $100–$500/month in credits for Max and Team subscribers encourages experimentation with agent development on the Claude Platform.

---

10. Future Directions

10.1 What to Watch

1. Browser use SDK maturity: The beta browser use and computer use SDKs will determine how easily developers can build production agent workflows on Haiku 5.5.

2. Effort setting optimization: As developers gain experience with adjustable effort, best practices for tuning cost-intelligence trade-offs will emerge.

3. Multi-agent architecture patterns: The combination of Opus 5.5 + Sonnet 5.5 + Haiku 5.5 enables sophisticated multi-tier agent systems — real-world patterns will evolve quickly.

4. Competitor response: GPT-6 Luna's matching $0.10/M pricing suggests the small-model price floor is established. Expect further differentiation on features rather than price.

5. Enterprise adoption: The combination of low cost, 1M context, and browser use could make Haiku 5.5 attractive for enterprise automation — particularly for repetitive tasks like form filling, data entry, and document processing.

6. Claude Code integration depth: Haiku 5.5's availability in Claude Code could drive adoption among developers for simple coding tasks and code reviews.

7. Open-weight competition: The continued rise of open-weight models (Qwen3.8-27B, DeepSeek V4.1-Flash) could pressure Anthropic's closed-model economics in the small-model tier.

8. Haiku 6 development: If the 5.5 family follows the pattern of previous generations, a Haiku 6 release could bring further improvements in capability and efficiency.

---

11. References & Resources

Official Sources

• Claude Haiku 5.5 Announcement — Official launch post with benchmarks, pricing, and customer quotes
• Claude Haiku Product Page — Model overview, use cases, and availability
• Haiku 5.5 System Card — Full safety evaluation methodology and results (PDF)
• Migration Guide — Transitioning from Haiku 4.5 to Haiku 5.5
• Browser Use SDK — Computer use and browser use documentation
• Life Sciences Verification Program — Biology safeguards verification
• Cyber Verification Program — Cybersecurity safeguards verification

Related Research in This Journal

• Grok 4 7 Frontier Coding Knowledge Work New Safeguard Stack Longer Rl 2026 10 06
• Qwen3 8 Omni Flash Native Omnimodal Agentic Media 1m Context 2026 10 05
• Pareto 26 10 Preview Blended Model Pareto Frontier Cost Performance Unbiased 2026 10 02
• Claude Sonnet 5 5 Faster Cheaper Coding Jump Cyber Safeguards Distillation Protection 2026 09 29
• Claude Opus 5 5 Fable 5 1 Performance 40 Percent Cost Reduction Pacing Frontier Safeguards 2026 09 23
• Gpt 6 Sol Luna 50 Percent Price Cut Agentic Coding Cost Efficiency Frontier 2026 09 24

External References

• Artificial Analysis: Claude Haiku 5.5 — Independent benchmark scores and cost analysis
• Simon Willison: Claude Haiku 5.5 — Technical analysis and pricing breakdown
• BenchLM: Claude Haiku 5.5 — Benchmark aggregation and comparison

---

This article was researched using official Anthropic sources (Haiku 5.5 announcement page, Haiku product page, system card PDF, migration guide, and browser use SDK documentation). All benchmark figures are cited directly from Anthropic's published launch materials with appropriate caveats about vendor-reported results. Customer quotes are reproduced verbatim from the official announcement.