Anthropic Ships Claude Opus 5.5, OpenAI Counters with GPT-6 Sol and Luna Hours Later, Triggering Sharp Price Cuts

Two labs, one day, a new price floor

Anthropic released Claude Opus 5.5 on September 22, 2026. Roughly an hour later, OpenAI released GPT-6 Sol and GPT-6 Luna. The near-simultaneous launches follow Grok 4.7 and MiMo v2.6 Flash/Pro, both released the previous day, making it four notable model updates in 48 hours.

The headline number: GPT-6 Luna and GPT-6 Sol are priced at roughly half of their GPT-5.6 equivalents.

Pricing breakdown

Model	Input	Cached input	Output
GPT-6 Luna	$0.10/M	$0.01/M	$0.50/M
GPT-5.6 Luna	$0.20/M	$0.02/M	$1.20/M
Grok 4.7	$2/M	$0.50/M	$6/M
GPT-6 Sol	$2/M	$0.20/M	$10/M
GPT-5.6 Terra	$2/M	$0.20/M	$12/M
Claude Opus 5.5	$4/M	$0.20/M	$20/M
GPT-5.6 Sol	$4/M	$0.40/M	$20/M
Claude Fable 5.1	$10/M	$0.25/M	$50/M
GPT-6 Astra	$10/M	$1/M	$50/M

GPT-5.6 pricing carries a scheduled 25% increase in November, meaning GPT-6 Sol and Luna are effectively half the price of the promotional GPT-5.6 rates that will soon expire. GPT-6 Sol now matches GPT-5.6 Terra's price point, effectively eliminating any reason to choose Terra going forward.

At $0.10/$0.50 per million tokens, GPT-6 Luna is among the cheapest models OpenAI has released, trailing only GPT-4.1 Nano ($0.10/$0.40, April 2025) and GPT-5 Nano ($0.05/$0.40, August 2025) in raw price — while reportedly offering substantially stronger performance.

Claude Opus 5.5 also got cheaper. Every Opus release from 4.5 through 5.0 held steady at $5/M input and $25/M output. Opus 5.5 drops that to $4/M and $20/M — a 20% cut — with cached input pricing falling 60%, from $0.50/M implied under the old rate structure down to $0.20/M. Anthropic says Sonnet 5.5 and Haiku 5.5 are coming soon; current Haiku 4.5 pricing ($1/$5) is ten times more expensive than GPT-6 Luna, putting pressure on Anthropic's budget tier.

Communication style and behavior changes

According to Anthropic engineer Thariq Shihipar, Opus 5.5 is designed to address complaints about the model's communication style, claiming it now matches the intelligence of Claude Fable 5.1 while being more token-efficient "across every effort level." Anthropic also claims improved performance on Blender-based tasks, according to engineer Alex Albert.

Early independent testing surfaced a notable failure mode: developer Simon Willison's standard "pelican riding a bicycle" SVG benchmark caused Opus 5.5 running at "max" thinking effort to exhaust its 128,000-token output limit mid-reasoning, twice, without producing a final image. The model's internal reasoning trace showed it deliberating over minute anatomical details — shin length, chainring teeth, foot placement — well past the point of diminishing returns. This suggests the max thinking setting may over-allocate reasoning tokens on simple tasks, a behavior that has not yet been verified across more complex workloads.

Context window sizes, parameter counts, and formal benchmark scores for all three models have not been disclosed as of publication.

What this means

The roughly 50% price cuts from OpenAI and 20% cut from Anthropic, landing within an hour of each other, indicate the frontier-model price war has intensified rather than stabilized. Grok 4.7's earlier positioning at $2/$6 — undercutting GPT-5.6 Sol by more than half — has already been matched or beaten by GPT-6 Sol's $2/$10 pricing. Mid-tier models are now the primary battleground; flagship-tier pricing (GPT-6 Astra, Claude Fable 5.1, both at $10/$50) has held steady, suggesting labs are competing hardest on cost-efficiency for high-volume, agentic use cases rather than absolute peak capability. The Opus 5.5 max-effort failure is a reminder that lower prices and higher claimed intelligence don't eliminate reliability questions — output token limits remain a hard constraint that reasoning models can still hit unexpectedly.