GPT-6 Sol, Luna & Claude Opus 5.5 Released: Price Drops, Benchmarks & How to Choose

Three Major Model Releases in One Night

On September 23 (Beijing time), Anthropic released Claude Opus 5.5 and roughly 100 minutes later OpenAI added GPT-6 Sol and GPT-6 Luna to the GPT-6 family. Both vendors highlight stronger, faster, or more cost‑efficient models, with pricing reductions featured prominently.

GPT-6 Sol and Luna: New Low‑Price Tiers

OpenAI positions gpt-6-sol for complex coding and agent workflows, and gpt-6-luna for budget‑sensitive, high‑volume tasks; the existing gpt-6-astra continues to serve the most demanding workloads.

Price cuts: Sol’s regular input and output prices are halved versus GPT‑5.6 Sol. Luna’s input is halved and output drops from $1.20 to $0.50 per million tokens (≈58% reduction).

20× price gap: Sol and Luna differ by a factor of 20 in per‑token pricing, making Luna a viable candidate for classification, extraction, and format‑conversion tasks.

AutomationBench chart: OpenAI’s own chart shows Sol at xhigh reasoning intensity scoring 33.2% at $0.27 per task, while Opus 5 at max scores 26.9% at roughly 11.1× the cost. The comparison uses Opus 5 (not 5.5), different tools/reasoning settings, and cost‑per‑task (not cost‑per‑successful‑delivery).

Reduced misleading completions: OpenAI reports fewer cases where the model claims a coding task is done when it is not, based on a targeted “deception” evaluation.

Availability: Plus, Pro, Business, Enterprise, and Edu users can access Sol/Luna in ChatGPT Work and Codex; Free and Go users get Luna in the desktop app. API model names are gpt-6-sol and gpt-6-luna.

Claude Opus 5.5: Capability, Speed, and Writing Style Improvements

Anthropic’s first Claude 5.5 family model, Opus 5.5 , targets agent coding, computer operation, and knowledge work. Sonnet 5.5 and Haiku 5.5 are slated for later release.

Benchmark claims: Terminal‑Bench 4.0: Opus 5.5 66.4% ( xhigh ) vs Opus 5 52.3%; FrontierCode v1.1: 54.4% vs 48.0% (both at max reasoning). Different reasoning tiers are used, so direct column‑by‑column comparison is cautioned.

Default‑tier efficiency: Anthropic emphasizes that Opus 5.5 at default settings beats GPT‑6 Astra at roughly one‑fifth the per‑task cost on FrontierCode, and reaches similar Terminal‑Bench 4.0 performance at about 40% of the per‑attempt cost.

Writing style: Opus 5.5 presents key information first, uses less jargon, and adheres better to user‑specified writing rules. An official example shows Opus 5.5 explaining a billing discrepancy before diving into code.

Safety: Best score in Anthropic’s automated behavioral audit; pre‑release external evaluations by METR and Frontier Design (participation ≠ endorsement of all capability/cost claims).

Availability: On AWS, Google Cloud, Microsoft Azure, and Claude Platform (model name claude-opus-5-5). Pro, Max, Team, and per‑seat Enterprise plans receive increased 5‑hour quotas plus a one‑time quota reset.

Head‑to‑Head Pricing Comparison

Standard API prices (USD per million tokens):

GPT‑6 Sol: input $2, cache read $0.20, cache write $2.50, output $10.

GPT‑6 Luna: input $0.10, cache read $0.01, cache write $0.125, output $0.50.

Claude Opus 5.5: input $4, cache read $0.20, cache write $5, output $20.

Sol/Luna: requests exceeding 272k input tokens double input/cache prices and multiply output by 1.5×. Opus 5.5 regular input/output are 20% lower than Opus 5; cache read is 60% lower. Anthropic’s “typical workload cost down 40%” combines price cuts with fewer tokens used — not a guaranteed 40% bill reduction for every user.

Speed: Opus 5.5 output >30% faster; Fast mode (Claude Code/Platform) up to 2.5× speed but at $8/$40 per million tokens input/output.

Cache Economics: Hit Rate Matters More Than Unit Price

Both Sol and Opus 5.5 charge $0.20 per million tokens for cache reads, but cache write prices differ ($2.50 vs $5).

OpenAI’s “up to 90% cache discount” applies only to hit input tokens; output is billed normally, and the first cache write costs 1.25× regular input.

Worked example (Sol): 100k input / 10k output per turn, 80k fixed prefix cached. Turn 1: $0.340. Turn 2 (full hit, no new write): $0.156. Two‑turn total $0.496 vs $0.600 without cache ≈17% savings — not 90%. Cache prefix reusable for 30 minutes after last write/reuse.

Continuous edits around the same project benefit most; one‑off short requests may not justify cache engineering.

Independent Evaluation: Price Cuts ≠ Uniform Capability Gains

Artificial Analysis testing reveals nuance:

Intelligence Index (max tier): Sol per‑task cost fell from $1.99 to $1.06; Luna from $0.18 to $0.07. Output token usage slightly increased, so savings stem mainly from price reductions.

Coding Agent Index: Sol improved by 2 points; Luna regressed by 2 points. Some knowledge‑work tasks showed regressions (missed requirements).

Environment and task sets differ from vendor benchmarks.

Practical Recommendation: Test With Your Own Workload

Run a representative batch of your recurring tasks with the new models and your current model under identical acceptance criteria.

Simple classification/extraction → try Luna first.

Complex coding → include both Sol and Opus 5.5 in the candidate set.

Heavy repeated document/code reading → measure cache hit rates.

Track total spend (model calls, tool use, human verification) divided by accepted tasks. Count retries, model upgrades, and human cleanup in total cost; also monitor latency and unacceptable errors.

The real question: with the same budget, how many more qualified results can you deliver?