New Anthropic, OpenAI models make same promise: A little more for a lot less money

Anthropic and OpenAI both shipped efficiency-minded models: Anthropic’s Opus 5.5 as a cheaper, faster flagship for coding and knowledge work, and OpenAI’s GPT-6 Sol and Luna as mid and small tiers trained with methods akin to GPT-6 Astra. The pitch on both sides is the same — a little more capability for a lot less money.

That lands as enterprises compare vendors, stand up model routers, and flirt with open-weight alternatives. The frontier race has entered its comparison-shopping phase.

We’re less interested in the leaderboard theater than in the invoice. Anthropic priced Opus 5.5 at $4 and $20 per million input and output tokens — 20% below Opus 5 — with cache reads at $0.20 per million, 60% cheaper, and output more than 30% faster. Anthropic also claims roughly 40% savings on typical default workloads because the model uses fewer tokens to finish tasks. Partner benchmarks show modest coding and knowledge-work edges over GPT-6 Astra in some cases. High-risk cyber and biology prompts still get the Fable 5.1-style transparent routing to older models.

OpenAI’s Sol and Luna story is similarly economic. Sol sits as the capable daily driver under Astra; Luna is the fast cheap tier. API prices: Sol at $2 / $10 per million input/output tokens; Luna at $0.10 / $0.50. Depending on the benchmark, OpenAI says they land a few points above predecessors at half the cost. Rough mappings put Astra near Fable, Sol near Opus, Terra near Sonnet, Luna near Haiku — imperfect, especially at the top.

Useful AI wins when good-enough models become affordable enough to run in production harnesses all day. That is the honest upside. The downside is subtler: price wars can paper over whether orchestration, evaluation, and organizational practice — not raw model IQ — are what actually deliver value. Ars Technica’s Samuel Axon is right that harnesses matter as much as weights. Cost relief may also become a quiet brake on demand for ever-larger frontiers while safety advocates argue for intentional slowdowns.

I’m watching whether Opus 5.5’s token thrift shows up in real agent bills, whether Sol displaces Astra spend for coding teams, and whether open-weight routers keep forcing these cuts. The trajectory if this becomes normal is a market that optimizes dollars-per-successful-task — healthier for buyers, harder for labs that need premium pricing to fund the next leap.

Ars Technica by Samuel Axon, Sep 22, 2026. Context includes GPT-6 Astra’s earlier September release and rising enterprise use of routers toward cheaper or open-weight models.