Does a cheaper small model make browser agents ready to scale?
Claude Haiku 5.5 sharply cuts Anthropic’s small-model token prices and posts large gains on the company’s selected computer-use tests. That makes high-volume automation more plausible—but the evidence is vendor-run and does not establish error-free performance in live workflows.
The short version
- What happened
- For prompts of up to 100,000 tokens, Anthropic cut Haiku input and output prices by 90% from Haiku 4.5; longer prompts are priced 50% lower. The company estimates an average saving of about 75%, after accounting for request mix and tokenizer changes.
- Who it affects
- Workers may spend less time on repetitive classification, summarisation and browser steps, but may also inherit more review work if automation quality is uneven.
- What remains unproven
- The central evidence comes from Anthropic, which developed and sells the model; the system card is disclosure, not independent certification.
Keep in mind
The central evidence comes from Anthropic, which developed and sells the model; the system card is disclosure, not independent certification.
Headline benchmark results are point estimates under selected settings, and the launch page does not provide confidence intervals for the main comparison table.
The OSWorld result uses an offline subset and should not be treated as the success rate for a live browser workflow.
Customer examples use different tasks, comparators and outcome measures; several lack a disclosed denominator on the release page.
List-token savings do not include integration, review, retries, observability, security controls or the cost of consequential errors.
Source analysis · 2 linked sources. Primary analysed source date: 7 October 2026.
Read the full report · 9 min The evidence, methods, implications and limitations
The direct answer: cheaper agents are more plausible, but not automatically dependable
Claude Haiku 5.5 changes the economics of using a small model for repetitive digital work. Anthropic launched it on 7 October as its fastest model at standard speed and priced requests of up to 100,000 tokens at $0.10 per million input tokens and $0.50 per million output tokens. Haiku 4.5 cost $1 and $5 respectively, so the new list prices are 90% lower in that prompt band. Above 100,000 tokens, the new rates rise to $0.50 for input and $2.50 for output, still half the previous model’s rates. Anthropic says roughly 90% of requests to Haiku 4.5 were in the lower band and estimates that a comparable workload will cost about 75% less on average after allowing for the new model’s tokenizer and output behaviour.
That is consequential for teams considering classification, summarisation, request routing, live support and browser automation. A task that was too expensive to run on every document or customer interaction may now fit a production budget, and a larger model can hand narrow subtasks to Haiku rather than doing all the work itself. But a price cut does not remove the cost of errors, review, retries, observability or incident response. The responsible conclusion is therefore conditional: Haiku 5.5 makes high-volume agents easier to justify economically, while each deployment still needs evidence that the whole workflow—not only the underlying model—meets its required accuracy and safety threshold.[1][2]
What Anthropic tested, and what the scores mean
The release compares Haiku 5.5 with Haiku 4.5, GPT-6 Luna and Anthropic’s larger Sonnet 5.5 across a selected set of knowledge-work, reasoning, coding, visual and computer-use benchmarks. The most striking browser result is 72.4% on an offline subset of OSWorld 2.1, compared with 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna in Anthropic’s table. On Terminal-Bench 4.0, which evaluates multi-step command-line tasks, Haiku 5.5 scored 39.2%, below Sonnet 5.5’s reported 70.6% but above the comparison models listed. The release also reports 45.9% on Humanity’s Last Exam without tools and 57.4% with tools.
These numbers show a large change under Anthropic’s evaluation settings; they do not supply a universal success rate. The OSWorld figure is explicitly for an offline subset rather than the full live benchmark, and a benchmark task may not reproduce the websites, authentication, changing interfaces, policies or data quality of an organisation’s real systems. The release page provides point estimates but no confidence intervals for its headline table. Model comparisons can also depend on prompting, tool scaffolding, effort settings, token budgets and software versions. The system card is valuable disclosure of the developer’s methods and risks, yet it remains developer-run evidence rather than independent replication or certification.[1][2]
The customer tests are useful signals, not a common denominator
Anthropic includes early tests from several customers, but they do not form one comparable study. AlphaSense says it ran 400 document questions and measured a statistically significant improvement from 0.76 to 0.84 over Haiku 4.5. HubSpot reports 92.8% averaged over three runs on its simulated CRM portal suite, without disclosing the number or composition of tasks on the release page. Asana says task-completion latency fell by more than 30% and inference per agent turn was up to 2.5 times faster than the model it currently uses. Box reports an 11-point gain at about half the latency, but does not provide a denominator or uncertainty interval on the launch page.
Those evaluations matter because they involve workflows closer to production than a general benchmark. They also illustrate why buyers need their own tests: each company measured a different task, comparator and outcome. The results were selected for Anthropic’s announcement and are presented as customer statements, not peer-reviewed studies or independently audited case reports. Before switching a live system, a team should define a representative test set, record false positives and false negatives separately, measure tail latency rather than only averages, and compare the total cost of successful completion—including retries and review—with the current process.[1]
Where the lower price could change work
The immediate opportunity is not an unsupervised general-purpose employee. It is a layer of narrow, high-frequency assistance: sorting incoming requests, extracting fields from known document types, condensing conversation history, drafting routine summaries, checking whether records need human attention or completing reversible browser steps. Anthropic has also added effort controls to the Haiku class, allowing developers to trade additional computation for better results on a task. Used carefully, that can make a system spend more on ambiguous cases and less on repetitive ones instead of applying the same model setting to every request.
For workers and customers, faster responses can reduce waiting and manual copy-and-paste. The distribution of risk matters, however. If a low-cost model makes automation cheap enough to apply everywhere, weak processes can scale as quickly as useful ones. Customer-service classifications can affect access to help; browser agents can submit, delete or purchase; document summaries can omit qualifying language. A human review step is not meaningful unless reviewers can see the source, understand what the model changed and stop or reverse the action. Organisations should also measure whether savings are shared through better service and less repetitive work rather than simply expanding monitoring or workload.[1]
Safety claims are stronger than before—and still self-assessed
Anthropic says Haiku 5.5 showed fewer misaligned behaviours and less willingness to cooperate with misuse than Haiku 4.5 across almost all of its alignment evaluations. The company also says the new model’s cyber safeguards are more restrictive than Haiku 4.5’s, although less restrictive than safeguards on some recent larger models. The release describes a deliberate boundary: defensive cybersecurity requests are more broadly permitted, while penetration testing and techniques Anthropic judges more likely to be used by attackers are blocked. Biology safeguards are said to match those used for Sonnet 5, Sonnet 5.5 and Opus 5, with broader access available through verification programmes.
Those controls are relevant because lower price and faster browser use can increase the number of consequential actions a system attempts. They are not evidence that safeguards will catch every harmful sequence, prompt injection or misuse pattern. Anthropic developed the model, selected the evaluations and operates the access controls; the system card is therefore transparency from an interested party, not an independent safety approval. Deployers still need least-privilege credentials, tool allowlists, confirmation for irreversible actions, logs, anomaly detection and a tested shutdown path. External red-team results and incident data would materially strengthen the assessment.[1][2]
How to decide whether the model is actually cheaper
A useful pilot should compare end-to-end outcomes, not multiply list prices by estimated tokens. Start with a bounded workflow and a versioned sample of real tasks stripped of unnecessary personal data. Measure task success, severe-error rate, review time, retries, latency at the slow end, tokens consumed and the share escalated to a larger model or person. Run the incumbent process and Haiku 5.5 under the same acceptance criteria. For browser work, include changed page layouts, interrupted sessions, ambiguous instructions and prompt-injection content, then verify that the agent fails safely rather than merely stopping less often.
This assessment would become more favourable if independent evaluators reproduce the reported computer-use gains, organisations publish denominators and failure categories from representative deployments, and lower token bills survive after review and fallback costs are included. It would become less favourable if live websites expose a wide gap from the offline subset, if safeguards are easily bypassed through multi-step tool use, or if organisations expand automation without monitoring rare but costly failures. Haiku 5.5 is available through Anthropic and major cloud platforms now, so those questions can be tested. The launch makes scale cheaper; evidence from scale must show whether it is also safe and useful.[1][2]