Claude Haiku 5.5 Review 2026: Benchmarks, Price and Speed

What Claude Haiku 5.5 is, and why Anthropic launched it now

Claude Haiku 5.5 arrived on October 7, 2026, completing the Claude 5.5 family after Opus 5.5 and Sonnet 5.5. Anthropic calls it the cheapest, fastest, and most capable small model it has ever released, built for high-volume, cost-sensitive work: summaries, classification, database queries, context compaction, live support, voice agents, and browser use.

The timing is not subtle. Reuters reports the launch comes as Anthropic expands its lineup ahead of a planned IPO, and it lands two weeks after OpenAI's GPT-6 Luna claimed the same $0.10/$0.50 price tier. The small-model segment has become the industry's price war, and Haiku 5.5 is Anthropic's counterpunch.

The Artificial Analysis numbers: intelligence per dollar

Artificial Analysis, the independent benchmark outfit, had Haiku 5.5 indexed on launch day across its effort settings. The Intelligence Index runs from 29 at low effort to 43 at max effort, against a median of 13 for reasoning models in the same price tier. Even the cheapest setting scores more than double the tier median.

Speed is where the effort dial shows its cost. Low effort generates 181 tokens a second with a 6.3-second time to first token; max effort reaches 243 tokens a second once it starts, but takes over 400 seconds to produce the first token. Max effort is a batch-compute mode, not a chat mode.

• The high-effort setting scores 38 on the Index; all five settings sit well above the tier median of 13.
• At a blended rate (7:2:1 cache-hit/input/output), Artificial Analysis puts the cost at $0.08 per million tokens.

Anthropic's own benchmarks: the honest read

Anthropic's launch figures show the biggest gains where Haiku used to be weakest. On the OSWorld 2.1 computer-use test it scores 72.4%, up from 15.7% for Haiku 4.5. On Terminal-Bench 4.0, an agentic coding test, it reaches 39.2% where its predecessor scored zero. FrontierCode 1.1 comes in at 46.4%, and Humanity's Last Exam at 45.9% without tools, rising to 57.4% with them.

Vendor benchmarks deserve the standard discount: they run in the vendor's own harness. The useful signal is the independent one, and Artificial Analysis confirmed the direction on launch day. Read the exact percentages as evidence, not promises.

The price cut, with the fine print

For prompts under 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens, with cache reads at $0.01. Anthropic says about 90% of Haiku 4.5 requests fell under that threshold, which is why it can claim roughly 90% lower costs for typical use and about 75% lower on average.

The fine print matters. Above 100,000 tokens the price quintuples to $0.50/$2.50. And the new tokenizer counts roughly 30% more tokens for the same text, according to third-party analysis of the launch, so a 90%-cheaper headline shrinks once you re-estimate with the new token counts. Prompt caching (up to 90% off cached input, per Anthropic) and batch processing (50% off) are where the real savings compound.

Haiku 5.5 vs GPT-6 Luna vs the rest of the small-model tier

The head-to-head that matters is GPT-6 Luna, OpenAI's small model launched September 22, 2026 at the identical $0.10/$0.50 tier. On Artificial Analysis's Intelligence Index, Haiku 5.5 leads at both ends: 29 vs 22 at low effort, 43 vs 38 at max effort. On the OSWorld computer-use test, Anthropic reports 72.4% for Haiku 5.5 against 48.9% reported for GPT-6 Luna.

Luna's advantage is distribution: it already reaches free-tier users through ChatGPT's desktop app, while Haiku 5.5 meets consumers through Claude.ai subscriptions. For developers the APIs are priced identically, so the choice comes down to benchmark profile and ecosystem, not price.

• Mistral's ML4 sits at $0.68/$2.09 per million tokens, per industry reporting: roughly seven times Haiku 5.5's input price.
• Open-weights models average $0.83 per million tokens on OpenRouter, per the same analysis. Haiku 5.5 undercuts the open average while beating its tier's intelligence median.

Who Haiku 5.5 is for

• Teams running classification, extraction, summarization, or support at high volume, where per-call cost dominates.
• Agent builders who want a cheap, fast worker model under a stronger planner like Sonnet 5.5 or Opus 5.5.
• Anyone who burned money on Haiku 4.5: the same workloads cost roughly three-quarters less, per Anthropic's math.
• Not for you if your prompts routinely exceed 100K tokens, if you need the strongest reasoning available, or if you want a consumer product with a signup bonus.

What we still do not know

• There is no independent track record: launch-day benchmarks are the entire evidence base so far.
• Real-world latency and reliability at scale are unproven; Artificial Analysis's figures are single-endpoint measurements.
• The ~30% tokenizer inflation figure comes from third-party analysis, not Anthropic. Verify it against your own workloads.
• Anthropic's 75%-cheaper average is a vendor estimate built on its own usage mix.
• Long-term pricing is not promised; the 100K-token tier boundary could move.