Claude Haiku 5.5 vs Sonnet 5.5
Last updated: October 8, 2026
Haiku 5.5 and Sonnet 5.5 arrived the same day — October 7, 2026 — and Anthropic does not present them as competitors. Its release positions Sonnet 5.5 (with Opus 5.5) for complex agentic coding, and Haiku 5.5 for high-frequency, bounded work — explicitly including the role of coding subagent alongside the larger models.[1]
So the question is scope, not rank. Below: the at-a-glance numbers, Anthropic's own when-to-use-which guidance, and the cost math — including Sonnet 5.5's same-day cache read cut.
The verdict in 30 seconds
Narrow, high-frequency work
Haiku 5.5. Summaries, compaction, classification, database queries, live support — at $0.10 input / $0.50 output per million tokens, these jobs belong in the cheap tier.
Complex agentic coding
Sonnet 5.5. Anthropic casts it (with Opus 5.5) as the driver, and Haiku 5.5 as the subagent. Terminal-Bench 4.0: 70.6% vs 39.2%.
A full agentic system
Use both. The release frames Haiku 5.5 as a coding subagent alongside Sonnet 5.5 and Opus 5.5 — a two-tier pattern, not an either/or.
At a glance
| Dimension | Claude Haiku 5.5 | Claude Sonnet 5.5 |
|---|---|---|
| Input price (per 1M tokens) | $0.10 (prompt ≤100k) / $0.50 (prompt >100k) | $2.00 |
| Output price (per 1M tokens) | $0.50 (prompt ≤100k) / $2.50 (prompt >100k) | $10.00 |
| Cache read price (per 1M tokens) | $0.01 (prompt ≤100k) / $0.05 (prompt >100k) | $0.10 — halved from $0.20 on Oct 7, 2026 |
| GDPval-AA v2.1 (knowledge work) | 1620 | 1840 |
| AA-Briefcase v1.1 | 1578 | 1824 |
| OSWorld 2.1 (computer use, offline subset) | 72.4% | 83.9% |
| Humanity's Last Exam (no tools) | 45.9% | 56.9% |
| Terminal-Bench 4.0 (agentic coding) | 39.2% | 70.6% |
| FrontierCode 1.1 (Main) | 46.4% | 52.1% (Xhigh) |
All figures from Anthropic's official release table, October 7, 2026.[1] Cache writes per 1M tokens: Haiku 5.5 $0.125 (≤100k) / $0.625 (>100k); Sonnet 5.5 $2.50. Sonnet 5.5's FrontierCode score is listed at the Xhigh effort setting; per-level scores for other settings are not published as numbers.
When to use which
The release voice on this is unusually direct, so here it is as a division of labor rather than a leaderboard:
- Haiku 5.5 — the high-frequency tier. Anthropic lists it for summaries, compactions, database queries, classification requests, live customer support, and browser use, plus the role of coding subagent alongside Opus 5.5 and Sonnet 5.5.[1]
- Sonnet 5.5 — the agentic work tier. Complex agentic coding is named as Sonnet 5.5 / Opus 5.5 territory. The published table backs the split: Sonnet 5.5 leads on every shared benchmark row, by the widest margin on Terminal-Bench 4.0 (70.6% vs 39.2%) and OSWorld 2.1 (83.9% vs 72.4%).
- The both-pattern — driver plus subagent. Haiku 5.5 is billed as the layer that runs alongside the bigger models, not instead of them. If you are building a full agent, the intended shape is a Sonnet-class driver delegating bounded calls to Haiku-class workers.
The cost picture
At the up-to-100k prompt tier, input is $0.10 for Haiku 5.5 against $2.00 for Sonnet 5.5 per million tokens, and output is $0.50 against $10.00. Cache reads: $0.01 against $0.10. Separately, Anthropic frames Haiku 5.5 as costing around 75% less to run than Haiku 4.5 — its own predecessor, not Sonnet.[1]
The same-day news cuts the other way for Sonnet: its cache reads were halved from $0.20 to $0.10 per million tokens on October 7, and Anthropic says that change makes agentic work around 20% cheaper to run. Caching-heavy agent loops are where that shows up. The tier separation itself is untouched — routing bounded, high-frequency calls to Haiku 5.5 keeps the expensive tier reserved for the work that actually needs it.
Not yet confirmed on the release page: context windows, rate limits, and real-world throughput for either model. Marked here as unpublished rather than estimated.
Where they are deployed today
Anthropic publishes vendor-reported results from early users, which are directionally useful for understanding the split. HubSpot reports 92.8% task success for CRM tasks on Haiku 5.5, and Asana reports a latency drop of more than 30% with reasoning around 2.5 times faster per agent turn. Box reports an 11-point improvement with roughly half the latency. These are vendor-published figures, not independent benchmarks.[1]
Frequently asked questions
Should I use Haiku 5.5 or Sonnet 5.5 for agentic coding?
It depends on the role, not the task category. Anthropic's release positions Sonnet 5.5 (and Opus 5.5) for complex agentic coding, and names Haiku 5.5 as a coding subagent alongside them — not as the driver. The published gap is real: Terminal-Bench 4.0 is 39.2% for Haiku 5.5 and 70.6% for Sonnet 5.5. If the model has to drive multi-step coding work, that is Sonnet territory; if it summarizes, compresses, or answers a bounded question inside a larger loop, Haiku 5.5 is the cost-efficient tier.
Did Sonnet 5.5 get a price cut?
Its list prices are unchanged: $2.00 input and $10.00 output per million tokens. What changed on October 7, 2026: cache reads were halved from $0.20 to $0.10 per million tokens, which Anthropic says makes agentic work around 20% cheaper to run. Cache-heavy agent loops benefit the most.
Can Haiku 5.5 replace Sonnet 5.5 entirely?
For narrow, high-frequency tasks — classification, extraction, summaries, compaction, database queries, live customer support — the release makes exactly that case, and the per-token difference is large. For complex agentic coding, frontier computer use, or long multi-step work, the published scores say no: Sonnet 5.5 leads on OSWorld (83.9% vs 72.4%), Terminal-Bench 4.0 (70.6% vs 39.2%), and Humanity's Last Exam with tools (64.5% vs 57.4%).
How close is Haiku 5.5 to Sonnet 5.5 on knowledge work?
Closer than the price gap suggests, but not equal. GDPval-AA v2.1 is 1620 vs 1840 and AA-Briefcase v1.1 is 1578 vs 1824, both favoring Sonnet 5.5 while Haiku 5.5 sits far above the previous small tier. The practical read: quality-sensitive knowledge work still belongs on Sonnet 5.5; volume work with a bounded definition of done is where Haiku 5.5 wins on cost per task.
Sources
- Anthropic — Claude Haiku 5.5 (official release page) Published October 7, 2026; includes the model comparison table, the Sonnet 5.5 cache-read change, and the CS-agent case studies cited below.
Every figure on this page comes from Anthropic's official Claude Haiku 5.5 release page, published October 7, 2026 — including the Sonnet 5.5 cache-read change announced the same day. VerdictApi did not use third-party estimates; anything the release does not state is labeled Not yet confirmed.
Last updated: October 8, 2026