Anthropic’s Claude Haiku 5.5 cuts token prices by 90% for requests below 100,000 tokens, making repetitive AI work a more attractive target for enterprise deployment. Yet the company estimates actual workload savings of approximately 75% against Haiku 4.5, rather than 90%. For businesses processing documents or running agents, that gap puts the focus on system design: request size, token usage and task routing determine the bill.
Match the Model to the Work Being Delegated
Anthropic positions Haiku as a lower-cost worker alongside Opus and Sonnet, which it continues to recommend for demanding coding work. Its target assignments include document summaries, information classification, database queries and smaller jobs delegated by more capable agents. Those are useful boundaries for an enterprise rollout: define the task before choosing the model.
Financial AI company Rogo provides a concrete example. A larger model could build a financial presentation and delegate retrieval of a revenue figure to Haiku. Alex Wang, who works in applied AI at Rogo, describes its accuracy as sufficient for that role, with speed and cost supporting frequent use.
That division suggests a practical architecture. Keep planning and difficult judgment with the stronger model, then pass bounded retrieval or processing tasks to the cheaper worker. The business case depends on whether the handoff preserves enough context to complete the assignment without repeated retries or escalation.
Request Size Changes the Cost Calculation
Anthropic prices Haiku 5.5 at $0.10 for every million input tokens and $0.50 for every million output tokens below the 100,000-token boundary. Above it, those charges rise to $0.50 and $2.50 respectively. Shorter requests receive the advertised 90% reduction against Haiku 4.5; longer ones receive a 50% reduction.
Caching follows the same tier structure. Per million tokens, reads cost $0.01 below the boundary and $0.05 above it, while writes cost $0.125 and $0.625. Haiku 4.5 charges $0.10 for reads and $1.25 for writes.
Anthropic says roughly 90% of requests handled by Haiku 4.5 fit the shorter category. Its approximately 75% workload-savings estimate also reflects a revised tokenizer, the component that converts content into billable units. That tokenizer consumes somewhat more tokens for equivalent work, so multiplying existing usage by the new headline rate would overstate the expected saving.
Two billing details remain unresolved in the launch information: treatment of a request containing exactly 100,000 tokens, and which tokens count toward that boundary. Teams operating near the cutoff should clarify both before setting budgets or routing rules. Crossing it means paying five times the lower-tier rate.
Compare Prices Without Assuming Equal Capability
OpenAI’s GPT-6 Luna matches Haiku’s four lower-tier charges, including cache reads and writes. Its higher pricing begins beyond 272,000 input tokens, giving the two products different cost boundaries even where their base rates match.
The wider comparison also needs qualification. These listed prices exclude tool charges, caching, batch discounts, regional premiums and negotiated terms. They compare billing rates, not equivalent performance.
| Product and applicable tier | Input: dollars per million tokens | Output: dollars per million tokens |
|---|---|---|
| Claude Haiku 5.5, below 100,000 tokens | $0.10 | $0.50 |
| GPT-6 Luna, base pricing; surcharge above 272,000 input tokens | $0.10 | $0.50 |
| Gemini 3.5 Flash-Lite, standard pricing | $0.30 | $2.50 |
| Claude Haiku 5.5, above 100,000 tokens | $0.50 | $2.50 |
| Gemini 3.8 Flash, promotional standard pricing through Dec. 31, 2026 | $0.75 | $3.75 |
| Grok 4.3, below 200,000 prompt tokens | $1.25 | $2.50 |
| Grok 4.7, below 200,000 prompt tokens | $2.00 | $6.00 |
| GPT-6.1 Sol, base standard pricing | $2.00 | $10.00 |
| Claude Sonnet 5.5, listed launch pricing | $2.00 | $10.00 |
For a production decision, measure the cost of a successfully completed assignment. A cheaper call can lose its advantage if it requires more tokens, repeated attempts or correction by a larger model. Conversely, a narrowly scoped task that succeeds consistently can make low rates valuable across a high-volume workflow.
Effort Settings Affect the Performance Available to Agents
Anthropic’s published evaluations show Haiku 5.5 improving on Haiku 4.5 and outperforming GPT-6 Luna on the selected tests below. Sonnet 5.5 remains ahead across those results. These are vendor-reported measurements, without independent verification.
| Test | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|
| GDPval-AA v2.1 | 1,620 | 735 | 1,437 | 1,840 |
| AA-Briefcase v1.1 | 1,578 | 614 | 1,336 | 1,824 |
| OSWorld 2.1, offline subset | 72.4% | 15.7% | 48.9% | 83.9% |
| Terminal-Bench 4.0 | 39.2% | 0.0% | 16.4% | 70.6% |
| FrontierCode 1.1, main evaluation | 46.4% | Not supplied | 42.4% | 52.1% |
GDPval-AA v2.1 and AA-Briefcase v1.1 measure knowledge work and report scores rather than percentages. OSWorld 2.1 assesses computer operation, while Terminal-Bench 4.0 covers agent-based coding. Anthropic labels Sonnet’s FrontierCode 1.1 result as High effort.
Haiku 5.5 also introduces configurable effort, with medium selected by default. Anthropic’s Terminal-Bench chart places the roughly 39% result at maximum effort; medium reaches approximately 20%. Developers evaluating the default configuration should therefore avoid treating the headline coding result as its expected performance.
A useful pilot would hold the task and acceptance criteria steady while testing effort levels. Record successful completion, latency and total token consumption together. That provides a basis for deciding whether a more intensive setting earns its place in the workflow.
Customer Results Offer Workload-Specific Evidence
Anthropic describes Haiku as its fastest model under standard operation, while acknowledging that Opus runs faster in Fast Mode. It provides no tokens-per-second measurement. Capacity planning therefore still requires testing against the intended workload.
Yashodha Bhavnani, Box’s VP of AI Products, reports an improvement of 11 points against Haiku 4.5 and roughly half the latency, although the scoring scale is unspecified. Aaron Vinh, a staff software engineer at Asana, reports a reduction exceeding 30% in task-completion latency and inference per agent turn up to 2.5 times faster than an unnamed comparison model.
Ze’ev Klapow, a distinguished software engineer at HubSpot, reports a 92.8% average over three runs of its CRM evaluation, leading the models it tested. These customer accounts support testing Haiku in specific applications. They don’t establish comparable throughput across providers or predict the result for a different production stack.
Integrate Through Existing Platforms, Then Verify Restrictions
Anthropic says deployment is available through its platform as well as AWS, Google Cloud and Microsoft Azure. The API identifier is claude-haiku-5-5. Updates to the Python and TypeScript SDKs introduce beta functionality for browser and computer operation.
That distribution gives enterprise teams several integration paths, but platform availability alone doesn’t settle deployment readiness. Browser and computer capabilities remain beta. Teams should validate the relevant interfaces and task behavior before connecting them to business processes.
Security policy also changes the usable workload. Compared with Haiku 4.5, the new model applies tighter cybersecurity safeguards, including blocking penetration testing under standard controls. Organizations needing expanded cybersecurity or biology access can seek approval through Anthropic’s verification programs. A migration should check permitted use before assuming an existing workflow will transfer unchanged.
Caching and Credits Change the Wider Operating Budget
Anthropic is also reducing Sonnet 5.5 cache-read pricing from $0.20 to $0.10 for each million tokens. It estimates approximately 20% savings for typical agent workloads, with the actual benefit tied to how much stored context an application reuses. Some existing customers on Azure and Google Cloud will receive that reduction over the following days.
For a system that keeps complex work on Sonnet and delegates smaller assignments to Haiku, the two changes address different cost components. Cheaper delegation reduces the price of suitable tasks; lower cache-read charges reduce the expense of repeatedly using stored context. Both should be measured against the application’s actual request pattern.
Anthropic is rolling out monthly API credits during the launch week:
- Max 5x subscriptions receive $100.
- Max 20x subscriptions receive $200.
- Team subscriptions receive up to $500, shared across their users.
Those credits apply to any model accessed through Anthropic’s platform. Details on Team allocation and rollover are unspecified, so they should support pilot planning without being treated as a fully defined long-term budget allowance.
Haiku 5.5 makes it more practical to test a tiered AI architecture now: reserve larger models for demanding work and evaluate cheaper workers against clearly bounded assignments. The deployment decision should rest on completion cost, effort settings and policy fit. Businesses that establish those controls can assess the new pricing against real operating performance before expanding usage.
