Anthropic Launches Claude Haiku 5.5: 75% Cheaper, Same Price as GPT-6 Luna

One Month, Third Model

Opus 5.5 on September 22, Sonnet 5.5 on September 28, Haiku 5.5 on October 7. Anthropic assembled the 5.5 family’s three tiers in half a month.

Haiku is the family’s smallest, cheapest, fastest tier. Anthropic’s positioning is explicit: high-frequency, cost-sensitive tasks — classification, summarization, content extraction, real-time support, voice agents, in-app assistants. These scenarios run enormous call volumes; a penny saved times billions of calls is real money.

Pricing: From $0.10, Same as Luna

Two pricing tiers. Requests under 100K tokens: $0.10/MTok input, $0.50/MTok output; over 100K tokens: $0.50 input, $2.50 output — a straight 5x jump. Reuters reports the pricing is 75% cheaper than Haiku 4.5.

Lined up against the price sheets, the most interesting rival is OpenAI’s GPT-6 Luna. VentureBeat verified every vendor’s official October 7 price list: Luna’s four tiers match Haiku 5.5 exactly. By comparison, Google’s Gemini 3.5 Flash-Lite is $0.30/$2.50 and Grok 4.3 is $1.25/$2.50 — both pricier than Haiku 5.5.

Two caveats beyond sticker price. First, the price-jump thresholds differ: Haiku jumps 5x past 100K tokens; Luna’s threshold is 272K input tokens. Second, per-token price isn’t cost-per-task — how many tokens a model eats, how often it retries, how accurate it is, all move the bill. Haiku 5.5’s official benchmarks aren’t out yet; the independent tracker QCode notes “no official benchmark figures released.”

Two more ways to save: prompt caching saves up to 90%, batch processing 50%.

First Haiku with Built-In Cyber Safeguards

Haiku 5.5 is the first Haiku with built-in safeguards for a narrow class of high-risk cybersecurity requests; per the official line, most everyday tasks are unaffected.

It extends the 5.5 family’s running theme: Opus 5.5 claims to cut sandbox-escape attempts by 85% and added anti-distillation preserved thinking; the Cyber Verification Program, refreshed October 6, opens the strongest models to vetted white-hat hackers for bug hunting.

Where to Get It

All Claude.ai tiers: Free, Pro, Max, Team, Enterprise — web, iOS, Android. Developers use the Claude Platform API with model ID claude-haiku-5-5; it’s on AWS, Google Cloud, and Microsoft Foundry, and selectable in Claude Code.

What It Means for You

Anthropic’s moves this past month are clear: Opus 5.5 cut 40% to contest the flagship tier, Sonnet 5.5 filled the middle, Haiku 5.5 cut 75% to hold the low end. Every step squeezes price; the architecture barely moved. Reuters frames the burst as “rounding out the lineup ahead of a planned IPO” — three tiers complete, product-line story whole.

If you’re building high-frequency subtasks inside agent apps — classification, extraction, routing — there’s now a Luna-priced option. Official benchmarks aren’t out; check real scores and token consumption before committing.