← The Polylog AI Intelligence Brief
Morning Edition · Friday, October 9, 2026Published at 2:20 AM EDT · New York
Anthropic Prices Claude Haiku 5.5 at $0.10 per Million Input Tokens, Matching OpenAI's GPT-6 Luna
The cut is 90% from Haiku 4.5's list price, and it leaves Anthropic and OpenAI at identical rates in both the small-model and mid-tier brackets.
Anthropic released Claude Haiku 5.5 on October 7 and called it its fastest and cheapest small model, built for high-volume work. Prompts up to 100,000 tokens cost $0.10 per million input tokens and $0.50 per million output tokens. Longer prompts cost $0.50 and $2.50. That is a tenth of Haiku 4.5's rates. The model accepts text and images and has a 1-million-token context window. Cache reads cost $0.01 per million tokens. A Russian-language AI channel described the target workloads as bulk request processing, classification, automation, simple agents and API traffic.
The prices are identical to OpenAI's. GPT-6 Luna lists at $0.10 and $0.50, and GPT-6 Sol lists at $2 and $10. Claude Sonnet 5.5 launched on September 28 at Sonnet 5's list price of $2 and $10. The two labs now charge the same list prices at two of their three tiers. Anthropic is competing on throughput and capability per dollar at those prices. At the top tier, it says Opus 5.5 performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.
The capability claims come from Anthropic and from aggregators. On OSWorld, Haiku 5.5 reportedly scores 72.4% against 15.7% for Haiku 4.5. The low Haiku 4.5 figure suggests a harder, newer version of the benchmark, so readers should not compare it with older OSWorld numbers. Artificial Analysis gives Haiku 5.5 an Intelligence Index score of 43 at maximum effort and 34 at default effort, against 17 for Haiku 4.5. On Terminal-Bench 4.0 the model scores 39.2%. On the same benchmark, Sonnet 5.5 scores 70.6% and Opus 5.5 scores 66.4%, so the small model trails the mid-tier model by a wide margin on agentic coding.
For buyers, the routing decision now matters more than the choice of vendor. In an OpenAI case study, the legal-tech company LegalOn cut estimated daily Codex costs by 65% by assigning tasks to GPT-6 Astra, Sol or Luna according to difficulty and by managing budgets. This is a vendor-published figure without independent audit, but the mechanism is straightforward. When the cheap tier can handle computer use and classification, most traffic moves down a tier.
What this means
Identical list prices at two tiers mean the two leading closed labs have stopped competing on headline price in those brackets and now compete on capability per token, latency and cache economics. This compresses gross margin on high-volume traffic for both. The groups that benefit are application companies whose costs are dominated by inference (coding tools, customer support, document processing), because their cost of goods falls with every model generation. The groups exposed are open-weight hosting providers whose main selling point was price, since a $0.10 closed model with a 1-million-token context weakens the reason to self-host small open models. Private-market valuations of the labs assume that revenue growth outpaces this per-token decline, and that assumption is now being tested openly. The likelier outcome is that within weeks Google prices its next Flash-class model at or below $0.10 per million input tokens and the low tier settles near that level. The condition is that labs keep prioritizing agentic volume over near-term margin. A move by either OpenAI or Anthropic to raise small-model prices, or to ration capacity, would show that compute supply, not competition, sets the price.
Part of a tracked trend
Frontier Model Price War
Frontier API vendors increasingly compete on strategic price-cutting rather than pure capability, repeatedly launching or repricing models below prevailing rates to grab share and compressing industry inference margins; expect recurring below-rival pricing moves.
Townsquare
Start a discussion in Townsquare.
More from this edition
- OpenAI Withdraws Three of 722 AI-Written Math Papers as Mathematicians' Group Urges Boycott
- Nvidia and Microsoft Open Preorders for RTX Spark Laptops With Up to 128 GB of Unified Memory for Local Agents
- Two Security Papers Find Agent Defenses Miss Attacks: Skill Scanners Bypassed 94% to 100% of the Time, Triage Agents Miss 40% of Alerts
- OpenAI Bans Russian and Iranian Influence Networks That Used Fake Journalists and a Latin American Think Tank
- Anthropic's New Usage Policy Bars Claude From Recommending Whom Police Should Investigate or Arrest
- New Statistical Certificate Shows Separated Leaderboard Rankings Withstand About 2,000 Forged Votes
- Agent-Controlled Forgetting Cuts a Debugging Session's Prompt Tokens From 912,000 to 232,000 Without Retraining
Comments
0No comments yet.