Anthropic Launches Claude Haiku 5.5 with ~75% Lower Running Costs
On October 8, Anthropic officially launched Claude Haiku 5.5, calling it its cheapest, fastest and most capable small model yet, with average running costs about 75% lower than Haiku 4.5. It is available on Claude Platform and Claude Code, with Sonnet cache-read prices also cut in half. It is the first Haiku model with adjustable effort settings and adaptive thinking, directly affecting model selection for high-volume, cost-sensitive workloads such as customer service, batch processing and agents.
Confirmed
- Average running costs 75% lower than Haiku 4.5; tasks under 100K tokens are up to 10x cheaper (about 10% of the previous price), and Anthropic says such tasks account for 90% of prior Haiku request volume.
- Tiered pricing: for prompts under 100K tokens, $0.10 input / $0.50 output per million tokens, with cache reads at $0.01; above 100K tokens, $0.50 input / $2.50 output. Per Artificial Analysis, the price drops to 10% of the previous generation with an Intelligence Index score of 43, up 26 points in a year. The Prompt Engineering channel also cited an OSWorld score of 72% and noted better short-prompt pricing versus OpenAI's comparable small model.
- Anthropic says the model significantly improves over Haiku 4.5 in coding, computer use and knowledge work, and supports computer use and workflow tasks.
- First Haiku with an adjustable effort parameter, letting users trade off cost versus intelligence per task, plus adaptive thinking.
- Anthropic says the model shows major improvements in nearly all alignment evaluations with substantially reduced misbehavior rates.
- Positioned for high-volume, cost-sensitive tasks such as summarization and classification; works well as a coding sub-agent alongside Opus 5.5 and Sonnet 5.5, and is fast enough for real-time customer support and browser operation.
- Anthropic's Alex Albert noted Haiku 4.5 was released in 2025; Elon Musk rarely praised it publicly as "another good model." Developer bcherny called Haiku 5.5 a competent Haiku-class model; Lydia Hallie found it well suited for high-frequency, low-cost use cases.
Why it matters
- The 75% cost reduction dramatically lowers the barrier for small-model adoption, directly affecting high-volume cost-sensitive workloads and intensifying price-performance competition among small models.
- Gains in coding and computer use mean small models can take on tasks previously requiring larger ones.
- Adjustable effort and tiered billing give developers finer cost control: Lydia Hallie shared that API users can set the autocompact window to 100K to keep context in the cheaper billing tier, saving about 5x per million tokens.
2026-10-08 ~ 2026-10-09 · 59 related posts
- Episode 1: Claude Haiku 5.5 Leaked in Claude Code, Priced Against GPT-6 Luna(2026-10-08, 3 posts)
- Episode 2: Anthropic Launches Claude Haiku 5.5 with ~75% Lower Running Costs(2026-10-08, 59 posts)
- Episode 3: Claude Haiku 5.5 ranks 2nd in intelligence but real costs spiral(2026-10-08, 16 posts)
Primary sources
- [source] Anthropic launches Claude Haiku 5.5, claiming ~75% lower cost than Haiku 4.5 — claudeai · 2026-10-08
- Claude Haiku 5.5 brings major gains in coding, computer use, knowledge work — claudeai · 2026-10-08
- Haiku 5.5 targets high-volume tasks, works as a coding subagent with Opus/Sonnet — claudeai · 2026-10-08
- Haiku 5.5 is first Haiku with adjustable effort setting for cost-vs-intelligence — claudeai · 2026-10-08
- Haiku 5.5 best value under 100K tokens, which cover ~90% of prior Haiku traffic — claudeai · 2026-10-08
- Anthropic says Haiku 5.5 shows major alignment gains over Haiku 4.5 — claudeai · 2026-10-08
- Anthropic launches Claude Haiku 5.5: ~75% cheaper, 72.4% on OSWorld, Sonnet cache halved — claudeai · 2026-10-08
- Anthropic launches Claude Haiku 5.5: 75% cheaper than Haiku 4.5, Sonnet cache reads halved — ClaudeOfficial · 2026-10-08
- Anthropic Releases Claude Haiku 5.5 — AMBNNJ · 2026-10-08
- [source] Claude Haiku 5.5 pricing: $0.10/$0.50 per million tokens under 100K prompt — ClaudeDevs · 2026-10-08
- Anthropic quietly rolls out Claude Haiku 5.5, rollout now underway — mark_k · 2026-10-08
- Claude Haiku 5.5 launched: runs ~75% cheaper than Haiku 4.5 — lydiahallie · 2026-10-08
- [source] Set Haiku 5.5 autocompact to 100K to stay in the cheap $0.10/$0.50 pricing tier — lydiahallie · 2026-10-08
- Haiku 5.5 lands in Claude Code and Platform, running ~75% cheaper than Haiku 4.5 — brada · 2026-10-08
- Claude Haiku 5.5 ships with huge jumps: OSWorld 15.7%→72.4%, beats GPT-6 Luna across the board — mark_k · 2026-10-08
- Claude Haiku 5.5 Cuts Costs 75%, Scores 1620 vs 735 on GDPval, Adds Adjustable Effort — petrusenko_max · 2026-10-08
- Claude Haiku 5.5 rumored to bring big coding and computer-use gains over 4.5 — eyishazyer · 2026-10-08
- Anthropic ships Claude Haiku 5.5: +26 points in a year, 10x cheaper than predecessor — ArtificialAnlys · 2026-10-08
- Anthropic ships Haiku 5.5: major gains, 75% cheaper and faster than 4.5 — alexalbert__ · 2026-10-08
- Anthropic pricing page confirms Haiku 5.5 at $0.10 per million input tokens — RLanceMartin · 2026-10-08
- Musk praises Claude Haiku 4.5 as Anthropic's cheapest, fastest small model, ~75% cheaper — elonmusk · 2026-10-08
- Claude Haiku 5.5 matches GPT-6 Luna pricing but new tokenizer hides a 25% increase — RexDouglass · 2026-10-08
- Claude Haiku 5.5 launches Day 0 on Databricks with ~15% quality gain over 4.5 — matei_zaharia · 2026-10-08
- Anthropic launches Claude Haiku 5.5, ~75% cheaper to run than Haiku 4.5 — BLUECOW009 · 2026-10-08
- Claude Haiku 5.5 launches: beats GPT-6Luna on benchmarks, cuts price 90% under 100k tokens — 量子位 · 2026-10-08
- Anthropic Launches Claude Haiku 5.5: 75% Cheaper, Up to $500 Monthly API Credits — AI寒武纪 · 2026-10-08
24 near-duplicate retellings: AnthropicAI · mikeyk · daniel_mac8 · ClaudeDevs · rohanpaul_ai · rickasaurus · dotey · trq212 · gaganghotra_ · AGI Hunt · cedric_chee · bcherny · Prompt Engineering · daniel_mac8 · 赛博禅心 · almmaasoglu · mikegiannulis · BLUECOW009 · repligate · 新智元 · WorldofAI · Arindam_1729 · 智东西 · scaling01