This content was generated with the help of AI and published without prior human review. It may contain errors and is not professional advice. Report a problem: info@delionsoft.com.
Anthropic released Claude Opus 5.5 on 22 September, and OpenAI followed about an hour later with GPT-6 Sol and GPT-6 Luna — a day after Grok 4.7 and MiMo v2.6. The headline is not capability but cost: the new OpenAI models are priced at roughly half their GPT-5.6 equivalents, and Opus 5.5 came with its own reduction. The original report collects the current price table and some early hands-on impressions.
What happened
According to the original report, GPT-6 Luna is listed at $0.10 per million input tokens and $0.50 per million output, against $0.20/$1.20 for GPT-5.6 Luna. GPT-6 Sol sits at $2/$10, down from $4/$20 for GPT-5.6 Sol. Since GPT-5.6 has a 25% price increase scheduled for November, the effective gap against the older family will widen further. With GPT-5.6 Terra priced the same as GPT-6 Sol, the author sees no remaining reason to pick Terra.
Claude Opus 5.5 dropped from the long-standing $5/$25 to $4/$20, a 20% cut, and cache read prices fell by 60%. That last number matters more than it looks: in long agentic sessions most input tokens are billed at the cached rate, so the real cost of multi-step work falls faster than the headline figure. Anthropic has said Sonnet 5.5 and Haiku 5.5 are coming; the open question is whether Haiku, currently $1/$5, can compete at the low end against Luna's $0.10/$0.50.
One caution from the hands-on section: Opus 5.5 at the "max" thinking level failed twice on a simple SVG generation task, consuming the 128,000-token output limit while still reasoning. Each failed attempt cost $2.56 and took close to 20 minutes. The author concludes that the top effort level may be unreliable for real work.
Why it matters
For teams running LLM features in production, the unit economics of a product designed six months ago may no longer reflect reality. Features that were shelved as too expensive per request — document classification, ticket triage, search reranking, draft generation — move back into range at $0.10 per million input tokens. The cache read reduction changes the calculus for agentic workflows specifically, where repeated context dominates the bill.
The second lesson is about effort settings. Higher reasoning levels are not uniformly better; they can burn budget and time and still return nothing. Cost control in agentic systems needs hard limits, not just a model choice.
What to do
- Re-run your cost model with current prices before the November GPT-5.6 increase. If you are on an older tier at the same price as a newer one, the migration is arithmetic, not strategy.
- Test the cheap tier on your actual workload, not benchmarks. Cheap models handle a surprising share of routine tasks; keep the expensive model for the cases that measurably need it.
- Set output token ceilings, timeouts and per-request cost alarms. Treat a runaway reasoning loop as a failure mode you must handle, not an edge case.
- Keep your model layer swappable. Prices and rankings have changed several times this year; hardcoded vendor calls make each change a small project.