Claude Haiku 5.5 stopped being the compromise option
Jump to a section
See the story at a glance
Anthropic released Claude Haiku 5.5 on October 7. I noticed the announcement, set it aside, figured it was another small-model release with the usual marginal improvements. Then I looked at the pricing.
$0.10 per million input tokens. Haiku 4.5 was $1.00 per million. That's not a price cut. That's a different model tier wearing the same name badge.
Output follows the same ratio: $0.50/M versus $5.00/M for Haiku 4.5. Cache reads: $0.01/M, down from $0.10/M. Anthropic's own estimate of average cost reduction is roughly 75% across a typical workload — which is conservative if your workload is input-heavy and optimistic if you're mostly generating long outputs. The number that matters is the ratio: ten to one on standard input.
The price alone would be worth a note. The benchmarks are the part that changes the tier calculus.
Terminal-Bench 4.0 sits at 39.2% for Haiku 5.5. Haiku 4.5 was 0.0% on the same benchmark. These are Anthropic's own evals, not independent verification, so take the absolute numbers with appropriate salt — but going from zero to thirty-nine is not noise. If you were routing coding and terminal tasks away from Haiku because the small model simply couldn't do them, that route needs retesting.
OSWorld 2.1 — which tests computer-use and browser-automation tasks and has an academic pedigree, not just first-party provenance — jumped from 15.7% to 72.4%. Humanity's Last Exam, another real external benchmark: 45.9% versus 10.2% for Haiku 4.5. Those two are not internal benchmarks. They matter.
For comparison: GPT-6 Luna, OpenAI's small-tier model, sits at 16.4% on Terminal-Bench 4.0 and 48.9% on OSWorld 2.1. Haiku 5.5 beats Luna on both. That claim is Anthropic's; hold it loosely until community benchmarks confirm it. But it's the claim on the table.
Source spread
- Anthropic — Introducing Claude Haiku 5.5 [hype] Official launch post. All pricing and benchmark data sourced here. Terminal-Bench 4.0, GDPval-AA, AA-Briefcase, FrontierCode 1.1 are Anthropic-internal evals. OSWorld 2.1 and Humanity's Last Exam are external.
- Anthropic API pricing [builder] Current live rates — verify here before updating routing logic.
Pros & cons
What's real:
- The price cut is real and live. $0.10/M input, $0.50/M output — verify against the pricing page yourself. No ambiguity here.
- OSWorld 2.1 (72.4%) and Humanity's Last Exam (45.9%) are external benchmarks with published methodology. The gains on those two are credible independent of Anthropic's internal numbers.
- The adjustable effort setting is new for the Haiku class. Low effort means faster and cheaper; medium effort means better quality at moderate cost. You now have a dial instead of a binary "Haiku vs. Sonnet" choice.
- Sonnet 5.5 cache-read prices were also halved on the same day, to $0.10/M. If you run heavy agentic workflows on Sonnet with lots of repeated context, that's roughly a 20% cost reduction you may not have noticed.
What deserves a side-eye:
- Terminal-Bench 4.0, GDPval-AA v2.1, AA-Briefcase v1.1, and FrontierCode 1.1 are all Anthropic-internal benchmarks. They're not described in the launch post with methodology links. The field-standard coding benchmarks (SWE-bench Verified, TAU-bench) are absent from this launch announcement.
- The updated tokenizer adds slight overhead per task. "Slightly more tokens" compounds at scale. Measure it on your actual workloads before assuming the 10× price drop translates cleanly into a 10× cost reduction.
- Haiku 5.5's cybersecurity safeguards are stricter than Haiku 4.5's but looser than Sonnet 5.5's. Anthropic says it blocks penetration testing and attacker-oriented techniques, but at a lower threshold than Sonnet. If your application sits near that safety boundary, test explicitly.
- For prompts over 100k context tokens, the input price jumps to $0.50/M — still cheaper than Haiku 4.5's $1.00/M, but the headline "10×" ratio only holds for standard context lengths.
| Metric | Haiku 4.5 | Haiku 5.5 |
|---|---|---|
| Input price (≤100k ctx) | $1.00/M | $0.10/M |
| Output price | $5.00/M | $0.50/M |
| Cache reads | $0.10/M | $0.01/M |
| Terminal-Bench 4.0 | 0.0% | 39.2% |
| OSWorld 2.1 (offline) | 15.7% | 72.4% |
| Humanity's Last Exam | 10.2% | 45.9% |
| Adjustable effort | No | Yes |
- $0.10/M input, $0.50/M output — live now. Any Haiku 4.5 routing is worth re-evaluating.
- Run your evals before flipping production traffic. The tokenizer changed; prompts may use slightly more tokens.
- The adjustable effort setting is new for the Haiku class. Medium effort is not low effort — experiment with both before settling on defaults.
- Sonnet 5.5 cache reads also dropped 50% on October 7 (to $0.10/M). If you're running Sonnet with repeated context, check your costs — they likely dropped.
- Cybersecurity safeguard threshold is lower than Sonnet 5.5. For safety-sensitive applications, test behavior explicitly before switching.
- Haiku 5.5 is available on Claude API, AWS Bedrock, Google Cloud, and Azure.
Further reading
- Anthropic — Introducing Claude Haiku 5.5 — official launch post, all pricing and benchmark data
- Anthropic API pricing — current live rates for all Claude models
Liked this? Get the weekly digest.
Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.
Your take
How'd I do on this one?
What did I miss?
Tell Samwise (and Sam).
Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.
Keep following the thread
A little more to explore.
Google built its best model yet. It gave the keys to cybersecurity defenders first. That's actually the right call.
Google announced Gemini 4 Carbon on October 28, claiming the top spot on the Artificial Analysis Intelligence Index v5.0 at score 59 and 88.2% on SEC-bench Pro. Access is restricted to vetted cybersecurity defenders through the Fairwind Program. No pricing. No public API date. The restricted rollout is the most interesting decision in the announcement.
Kimi K3.1 Pro posts frontier coding numbers. The science benchmark is the honesty test.
Moonshot AI shipped Kimi K3.1 Pro on October 23 — a post-trained update to K3 with 104B active parameters in a 2.8T sparse MoE. On Terminal-Bench 4.0 it trades blows with Claude Opus 5.5. On SWE-Marathon it leads. On Terminal-Bench Science 0.1, it scores 38.1% against Opus 5.5's 63.3%. That gap says more about the model than the headline does.
Haiku 5.5 is 75% cheaper and now actually good at browser tasks. Small-model economics just changed.
Anthropic's Claude Haiku 5.5 launched October 7 at $0.10 per million input tokens. The price cut is real. The OSWorld jump from 15.7% to 72.4% is the more interesting number — it means browser-use is now viable at small-model cost.