Claude Haiku 5.5 vs GPT-6 Luna
They cost the same on paper — which small model is actually cheaper and better for your workload?
What will Claude or GPT actually cost per task?
Per-token prices hide the real bill: reasoning verbosity, cache reads, and long-context cliffs. Every price and benchmark here is source-linked and dated, and we only compare scores measured in the same run.
Independent & unaffiliated with OpenAI/Anthropic · pricing verified 2026-10-09 · how we source data
Claude cost decisions
The comparisons where pricing structure — not marketing — decides: cliffs, cache, and independent same-harness quality data.
GPT-6.1 Sol vs Claude Opus 5.5
Which model is better for coding and AI agents?
Claude Haiku 5.5 vs Claude Sonnet 5.5
Sonnet costs 20x more per token — when is it actually worth it?
Compare by workload
Start from what you're building. Each card links to the comparison that covers that scenario.
Model cost calculator
Real API cost estimates, not list-price trivia: service tiers (batch/flex/fast/ultrafast), per-request long-context banding, cached reads and cache writes, monthly volume — for Claude Haiku, Sonnet, and Opus 5.5 and the GPT-6 family. Combinations the vendor hasn't priced show Unsupported, never a guess.
Calculate your workload cost →Already have a usage export? Audit it against official rates · All tools
- Coding agent on a mid-size repo: Agent loop reading ~12K tokens of repo context per request, ~1/3 of it re-served from prompt cache on later turns. Typical for Claude Code / Codex-style tools.
- Long-document Q&A (crosses thresholds): One large document per request: 300K input tokens (100K cached from prior questions) — above OpenAI's 272K doubling threshold AND far above Haiku 5.5's 100K 5x cliff; only the flat-priced Claude models (Opus 5.5 and Sonnet 5.5) stay flat all the way.
- High-volume support chat: Short exchanges where most of the system prompt + history is cached. A 'good enough quality' workload where GPT-6 Luna is often the right pick.
Model lineup & standard prices
Standard-tier list prices per 1M tokens; click through to the calculator for tier, cache, and long-context rates.
| Model | Input | Cached | Output | Released |
|---|---|---|---|---|
| Claude Haiku 5.5(Anthropic)For high-volume, latency-sensitive tasks such as classification, extraction, and routing (Anthropic docs). | $0.10 | $0.01 | $0.50 | 2026-10-07 |
| Claude Opus 5.5(Anthropic) | $4 | $0.20 | $20 | 2026-09-22 |
| Claude Sonnet 5.5(Anthropic)The best combination of speed and intelligence (Anthropic docs). | $2 | $0.10 | $10 | 2026-09-28 |
| GPT-6 Sol(OpenAI) | $2 | $0.20 | $10 | 2026-09-22 * |
| GPT-6 Astra(OpenAI)Our most capable model for the most demanding work (OpenAI catalog tagline). | $10 | $1 | $50 | 2026-09-03 * |
| GPT-6 Luna(OpenAI)Our most efficient model for focused, high-volume tasks (OpenAI catalog tagline). | $0.10 | $0.01 | $0.50 | 2026-09-22 * |
| GPT-6.1 Sol(OpenAI)Near-Astra performance for complex work at a lower cost (OpenAI catalog tagline). | $2 | $0.10 | $10 | 2026-09-29 |
USD per 1M tokens, standard tier. Anthropic · OpenAI · * release date from a secondary source · verified 2026-10-09 against OpenAI and Anthropic pricing.
Latest model updates
Source-linked official announcements only.
- Anthropic ships Claude Haiku 5.5 at GPT-6 Luna's exact price
$0.10/$0.50 per MTok below 100K input tokens per request (5x above), 90% cheaper than Haiku 4.5, 'fastest model to date' at standard speed. Anthropic's own table shows it ahead of GPT-6 Luna on all six listed benchmarks; Vals AI independently scores it 54.31% overall.
Source: anthropic.com
- OpenAI releases GPT-6.1 Sol (DevDay)
Same $2/$10 list price as GPT-6 Sol, near-Astra evals, cached input halved to $0.10/MTok, Ultrafast tier priced. Available in ChatGPT Work and Codex at launch.
Source: openai.com
- Anthropic answers with Claude Opus 5.5 the same day
Opus 5.5 lands at $4/$20 per MTok — 20% below Opus 5 — with a 1M-token context window and 128K max output.
Source: anthropic.com
How we compare models
Source-verified only
Every number links to the official page it came from, with the date we checked. Unverifiable claims don't render.
Shared-run rule
Benchmarks compare only when they share a documented run. Publishers mix their own evals with third-party and competitor-cited numbers (we record who published AND who ran each figure) — the same score can even differ across publications.
Tier-aware costs
Batch −50%, Fast 2×, Ultrafast 6×, long-context ≈2× above 272K input tokens per request, cache reads/writes billed where published.
Evidence-graded picks
Recommendations carry Measured / Inferred / Insufficient-evidence labels. 'We don't know yet' is a valid answer.