Claude Haiku 5.5 Beats GPT-6 Luna at the Same $0.10 Price. Past 100K Tokens, Every Token Costs 5x.

What Anthropic shipped

Anthropic released Claude Haiku 5.5 on October 7, 2026, as claude-haiku-5-5, and called it “the cheapest, fastest, and most capable small model we’ve ever released.” Haiku 4.5 came out on October 15, 2025, so this is the first new Haiku in almost a year. It runs on Anthropic’s API, AWS, Google Cloud and Azure.

It is a much bigger model on paper: a 1M-token context window (Haiku 4.5 had 200K), 128K output tokens, and adjustable reasoning effort from low to max, with medium as the default. It is the first Haiku with effort levels.

Sources: Anthropic’s launch post and VentureBeat (standard prices), ComputingForGeeks (batch prices). The 100K threshold is counted on input tokens.

Anthropic’s own summary: 90% cheaper than Haiku 4.5 per token under 100K, 50% cheaper above it, and about 75% cheaper on an average workload once you count a new tokenizer that splits the same text into more tokens.

Against GPT-6 Luna, it isn’t close

GPT-6 Luna is OpenAI’s smallest GPT-6 model, at the same $0.10 / $0.50, with the same $0.01 cache read. Anthropic’s launch table puts the two side by side.

Source: Anthropic’s launch post. Vendor-reported; Haiku 5.5 at max effort. Sonnet 5.5’s FrontierCode score is at high effort.

The biggest gap is computer use: 23.5 points over Luna on OSWorld 2.1, and 4.6 times Haiku 4.5’s score. Anthropic also shipped beta computer use and browser use in its Python and TypeScript SDKs with this release, so that is clearly the job it has in mind.

The independent numbers agree on the direction. On Vals.ai’s Terminal-Bench 4.0 leaderboard, Haiku 5.5 scores 36.36% to Luna’s 13.64%. Artificial Analysis measured 33% in its own harness and gives Haiku 5.5 an Intelligence Index of 43, ahead of Gemini 3.8 Flash at 41.

The 100K cliff

Haiku 5.5 is the only current Claude model with a long-context tier. Once the prompt goes over 100K tokens, every token in that request, not just the overflow, is billed at 5x: $0.50 input, $2.50 output, $0.05 cache read. Luna’s step comes much later, at 272K, and is smaller: 2x input and 1.5x output.

Our math: cost of one request with 2,000 output tokens and no caching, at list prices. Haiku 4.5’s column ignores that the new tokenizer makes Haiku 5.5’s token counts 11% to 36% higher for the same text.

Anthropic says about 90% of Haiku 4.5 requests stayed under 100K, which is why the headline price holds for most traffic. Agents are the exception. A coding or computer-use agent re-sends its whole history on every step, and a long session crosses 100K and stays there.

What it costs on a real agent benchmark

Vals.ai runs Terminal-Bench 4.0 with the same minimal agent for every model, no step or cost limit, and publishes the cost per test. Its October 26 leaderboard shows the cliff in practice.

Source: Vals.ai Terminal-Bench 4.0 leaderboard, updated October 26, 2026. Cost per solved task is our math: cost per test ÷ score.

Haiku 5.5 solves 2.7 times as many tasks as Luna, at 8.9 times the cost per test. Per solved task it costs 2.7 times as much as GPT-6.1 Sol, a model with 20 times its token price. GLM 5.3 Flash scores a point higher at 39% of Haiku’s cost per test.

Two things drive that bill. Vals ran Haiku at max effort with a 128K output cap, and Haiku used 45.5 million tokens per test, more than any model on this list except Sonnet 5.5. Vals doesn’t publish the split between input, cached and output tokens, so we can’t say how much of the $3.12 came from the 5x tier. The math still points there: 45.5 million tokens in one task means either more than 455 calls averaging under 100K each, or calls that average above it. An agent’s history grows with every step, so a good share of those calls likely landed above 100K.

Effort is the other dial

ComputingForGeeks ran Haiku 5.5 on 8 small Python tasks at every effort level. Low and medium solved all 8 for under half a cent. Max solved one fewer and cost 15 times as much as medium.

Source: ComputingForGeeks, one run per task. A small test, but the max-effort blow-up is hard to miss.

Vals.ai ran Haiku at max effort, its most expensive setting. Your agent doesn’t have to.

Fun facts

• Haiku 4.5 scored 0.0% on Terminal-Bench 4.0 in Anthropic’s own launch table. Haiku 5.5 scored 39.2%.
• Above 100K, Haiku 5.5 costs $0.50 / $2.50: exactly Haiku 4.5 on the Batch API, and exactly GPT-6.1 Terra on Flex.
• Haiku 5.5 and GPT-6 Luna match to the cent on input ($0.10), output ($0.50) and cache reads ($0.01).
• Haiku 5.5’s 72.4% on OSWorld 2.1 is 1.8 points below what OpenAI’s chart gives GPT-6.1 Terra at max effort, at a tenth of Terra’s token price. Different vendors ran those tests, so treat it as a rough comparison.
• At max effort, Haiku 5.5 wrote 132,499 output tokens for 8 small Python tasks, about 16 times what it needed at medium.
• The same text costs more tokens: the new tokenizer produces 1.11x (JSON) to 1.36x (English prose) the tokens of Haiku 4.5.
• Anthropic’s model names track poem length: a haiku is three lines, a sonnet fourteen, an opus a whole work. The prices follow the same order.

What this means if you build with small models

Pick Haiku 5.5 over Luna for quality. At the same base price it wins every published benchmark, by 23.5 points on computer use. For classification, extraction, short chats and single computer-use steps, it is the better $0.10 model.

Keep prompts under 100K. Compact or summarize the history before it crosses the line. Capping an agent’s context at around 95K keeps every request on the base rate.

Use low or medium effort. Medium is the default for a reason. Max multiplies tokens and time without a matching gain on everyday work.

For long contexts, compare Luna. Between 100K and 272K, Luna costs a fifth as much per token. If the task needs a big context and not much intelligence, that gap decides it.

Measure cost per solved task. On Vals.ai’s agent benchmark, the cheapest token price didn’t produce the cheapest solved task. Run a week of your own work on two models and compare the bill and the pass rate.

Where MegaBrain fits

Full disclosure: we sell an AI coding agent. MegaBrain’s terminal coding agent and editor extensions run GLM, Kimi, MiniMax, DeepSeek, Qwen, Grok and Auto on the paid plans, Claude, GPT and Gemini paid from your credits from Pro, and any model through your own API key or local models, free on every plan. So you can put two models on the same task and compare the bills yourself. See the plans and subscribe. MegaBrain is not affiliated with Anthropic, OpenAI, Vals.ai or any of the sites cited here.

Sources

All pages opened October 27, 2026.

• Anthropic — “Introducing Claude Haiku 5.5” (October 7, 2026: prices, the 100K tier, effort levels, benchmark table against Haiku 4.5, GPT-6 Luna and Sonnet 5.5)
• VentureBeat — “Anthropic launches Claude Haiku 5.5 with 90% API price reduction, matching GPT-6 Luna” (Carl Franzen, October 7, 2026)
• ComputingForGeeks — “Claude Haiku 5.5: Pricing, Benchmarks, Tested” (specs, batch and cache prices, tokenizer, effort and speed tests)
• Vals.ai — Terminal-Bench 4.0 leaderboard (score, cost, tokens and duration per test; updated October 26, 2026)
• Vals.ai — Claude Haiku 5.5 model page (settings used: max effort, 128K max output)
• OpenAI — GPT-6 Luna model page (price, 272K long-context rule, knowledge cutoff)
• OpenRouter — Claude Haiku 5.5 (endpoints and the 100K-token price override)