Claude Haiku 5.5 vs Haiku 4.5
Last updated: October 8, 2026
Haiku 5.5 is not a soft refresh of Haiku 4.5. Anthropic says it costs around 75% less to run — 90% less for prompts of up to 100k tokens — and it outscores its predecessor on every benchmark in the October 7, 2026 release table.[1] The loudest single line: on Terminal-Bench 4.0, Haiku 4.5 scored 0.0% and Haiku 5.5 scores 39.2%.
Below: the at-a-glance comparison, three clear upgrade calls, and the one detail — the tokenizer change — that should shape how you re-plan a budget.
The verdict in 30 seconds
High-volume small tasks
Upgrade. Cheaper on every published rate and ahead on every published benchmark, for exactly the jobs Haiku is built for.
Complex agentic coding
Do not switch down. Terminal-Bench-class work belongs on Sonnet 5.5 or Opus 5.5; Haiku 5.5's official role there is subagent, not driver.
Cost-capped fleets
Upgrade. If Haiku 4.5 was pinned for budget, 5.5 undercuts it and raises the ceiling — the cost is revalidation, not price.
At a glance
| Dimension | Claude Haiku 5.5 | Claude Haiku 4.5 |
|---|---|---|
| Input price (per 1M tokens) | $0.10 (prompt ≤100k) / $0.50 (prompt >100k) | $1.00 |
| Output price (per 1M tokens) | $0.50 (prompt ≤100k) / $2.50 (prompt >100k) | $5.00 |
| GDPval-AA v2.1 (knowledge work) | 1620 | 735 |
| AA-Briefcase v1.1 | 1578 | 614 |
| OSWorld 2.1 (computer use, offline subset) | 72.4% | 15.7% |
| Humanity's Last Exam (no tools) | 45.9% | 10.2% |
| Terminal-Bench 4.0 (agentic coding) | 39.2% | 0.0% |
| Adjustable effort setting | Yes — first Haiku-tier model with it | No |
All figures from Anthropic's official release table, October 7, 2026.[1] Cache rates also fall: reads drop from $0.10 to $0.01 and writes from $1.25 to $0.125 per million tokens at the ≤100k tier.
Should you upgrade? Three clear calls
High-frequency, low-complexity workloads: upgrade. Classification, extraction, database queries, summaries, compaction, live support triage — the exact jobs Anthropic positions Haiku 5.5 for. Every published benchmark goes up and every per-token rate goes down versus Haiku 4.5. There is no published dimension where 4.5 still wins. The real cost of switching is revalidation: the tokenizer changed, so re-baseline token counts and prompts before trusting your old cost model.
Complex agentic coding as your main loop: do not stay on any Haiku for it. In Anthropic's own framing, Terminal-Bench-class coding is Sonnet 5.5 / Opus 5.5 territory — the release positions Haiku 5.5 as a coding subagent alongside them, not as the driver. Haiku 5.5's 39.2% on Terminal-Bench 4.0 is a leap from 0.0%, but Sonnet 5.5 posts 70.6% on the same benchmark. If you need multi-step agentic coding, the move is up, not sideways.
Cost-capped inference at scale: upgrade. If Haiku 4.5 was pinned for budget, 5.5 undercuts it on every published rate — input $0.10 vs $1.00, output $0.50 vs $5.00, cache reads $0.01 vs $0.10, cache writes $0.125 vs $1.25 per million tokens at the ≤100k tier — while scoring higher everywhere in the table. Caching-heavy pipelines save the most.
Decision rule: upgrade by default. The only workload that should not move to Haiku 5.5 is one that needs frontier-level agentic coding — and that workload belongs on Sonnet 5.5 or Opus 5.5, not on Haiku 4.5 either.
Terminal-Bench 4.0: 0.0% vs 39.2%
No other row in the release table moves from zero. In the published run, Haiku 4.5 completed the benchmark's agentic coding tasks at a 0.0% pass rate; Haiku 5.5 lands at 39.2%. For scale: GPT-6 Luna sits at 16.4% in the same table, and Sonnet 5.5 at 70.6%.
Read it precisely. A 0.0% published score means the previous small tier was not usable for this class of multi-step coding work; 39.2% puts the small tier on the board for the first time — at a small-model price. It does not close the gap with Sonnet 5.5: the tier separation in coding is intact, which is exactly why Anthropic frames Haiku 5.5 as the subagent layer.
What changes on your side
- Model ID —
claude-haiku-5-5, a new model rather than a version tag on the old one. - Tokenizer — updated, and it produces slightly more tokens per task, per Anthropic. The published discount figures already account for this; your old token-count baselines do not.
- Pricing tiers — two prompt-length tiers (≤100k and >100k tokens) instead of one flat rate. Long-prompt pipelines bill differently from short-prompt ones.
- Still unpublished — context window, rate limits, and real-world throughput are not stated on the release page. Marked here as Not yet confirmed, not estimated.
- Haiku 4.5 status — no sunset announced; it stays listed with its existing prices.
Frequently asked questions
Is Haiku 5.5 a drop-in replacement for Haiku 4.5?
Plan for it as a model swap, not a config change: the model ID is new (claude-haiku-5-5), the tokenizer is updated and produces slightly more tokens per task, and pricing now has two prompt-length tiers. Re-run your evaluations and cost baselines rather than assuming parity.
Why is Haiku 4.5's Terminal-Bench 4.0 score 0.0%?
That is what Anthropic's release table lists, and the page does not explain it further. The practical reading: the previous small tier was not completing that benchmark's agentic coding tasks in the published run, and the new one does — Haiku 5.5 scores 39.2% on the same benchmark.
Did Anthropic discontinue Haiku 4.5?
No shutdown notice appears on the release page — Haiku 4.5 is still listed there with its existing prices. We found no verified sunset announcement. If Anthropic publishes one, this page will be updated rather than left to age.
Is Haiku 5.5 cheaper and better, or just cheaper?
On the published numbers, both. Every per-token rate is lower than Haiku 4.5's in the same category, and every benchmark in Anthropic's comparison table shows a higher score for Haiku 5.5. The caveats: the benchmarks are vendor-run, and the tokenizer change means cost per task should be re-measured on your own traffic.
What does the upgrade mean for my cost model?
Re-baseline it. Prompts of up to 100k tokens are 90% cheaper per Anthropic's footnote and prompts above 100k are 50% cheaper, with the updated tokenizer already factored into those figures. Your old token counts are not comparable, because Haiku 5.5 produces slightly more tokens per task.
Sources
- Anthropic — Claude Haiku 5.5 (official release page) Published October 7, 2026; includes the Haiku 4.5 comparison table.
Every figure on this page comes from Anthropic's official Claude Haiku 5.5 release page, published October 7, 2026. VerdictApi did not use third-party estimates; anything the release does not state is labeled Not yet confirmed.
Last updated: October 8, 2026