Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna all shipped yesterday — here's what the 40% price cut means
Anthropic and OpenAI both launched new models within an hour of each other yesterday, cutting prices 40-50% across the board. Here's what the simultaneous drops mean for production deployment economics.
Yesterday Anthropic announced Claude Opus 5.5 at 2:14 PM ET. OpenAI announced GPT-6 Sol and GPT-6 Luna 63 minutes later. Both companies cut their flagship model prices by roughly 40%.
Opus 5.5 runs at $9/M input tokens and $45/M output — down from Opus 5's $15/$75. OpenAI's Sol matches GPT-6 Astra's benchmark performance at $10/$30, half Astra's $20/$60. Luna sits below Sol at $4/$12, matching GPT-5.3 Turbo's capability at less than half the old $10/$30 rate.
The timing wasn't coordinated. Anthropic's blog post references OpenAI's June pricing but makes no mention of Sol or Luna. OpenAI's launch post ignores Anthropic entirely. Both companies clearly had these releases queued and neither blinked first.
What changed in 90 minutes
Before yesterday, deploying Opus 5 for a high-volume voice agent meant budgeting $75 per million output tokens. A 30-second response at 40 tokens/second cost roughly $0.09 in generation alone. Scale that to 10,000 calls per day and the model cost hit $900 daily, $27k monthly, before infrastructure or telco.
Opus 5.5 cuts that to $0.054 per 30-second response. Same volume drops to $540 daily, $16.2k monthly. For the workflows we deploy — inbound restaurant bookings, NEMT dispatch confirmation, retail order status — that's the difference between break-even at 3,000 calls/month and break-even at 1,800.
Sol's $30 output pricing undercuts Opus 5.5 by 33% if benchmarks hold. Luna at $12 makes GPT-class reasoning accessible to use cases that couldn't justify GPT-5.3 Turbo six months ago.
The benchmark picture is messier than the pricing
Artificial Analysis published a comparison within 4 hours of the announcements. Opus 5.5 scores 88.3% on their Quality Index vs. Fable 5.1's 88.9%. Sol sits at 87.1%, Luna at 84.6%. All four models cluster within a 4.3-point range, which is tighter than the gap between GPT-5.3 Turbo (81.2%) and GPT-6 Astra (87.8%) was in May.
The Decoder's early testing found "little gain in actual intelligence" from Sol and Luna compared to Astra. Opus 5.5 shows the same pattern — Anthropic's own launch post says it "matches Fable 5.1 on most tasks" rather than exceeds.
What we're seeing is a maturity plateau. The capability delta between adjacent model generations has compressed while the efficiency delta widened. Opus 5.5 delivers Fable 5.1 performance at 40% lower cost per token. Sol delivers Astra performance at 50% lower cost. These aren't intelligence leaps; they're deployment-economics shifts.
The on-premise question nobody's answering yet
Sol and Luna ship with the same API surface as Astra. Opus 5.5 ships with the same API surface as Opus 5. Neither company published updated hardware requirements for self-hosted deployments, and neither company's enterprise SKU page reflects yesterday's models yet.
For the multi-tenant CRMs we build, the question isn't just price-per-token. It's: can a client run Sol locally on the same 8×H100 rack that currently runs Astra, or does the efficiency gain only apply to API calls? Can Opus 5.5 fit in the same 4×A100 config that Opus 5 required, or did architectural changes blow the VRAM budget?
Those specs will ship eventually. Until they do, the 40% savings only apply to metered cloud usage, not to the self-hosted enterprise deployments where TCO modeling depends on hardware amortization over 36 months.
What this means for October roadmaps
We have three client deployments slated to move from GPT-5.3 Turbo to Astra in October. All three justified the 2× cost increase based on Astra's improved multi-turn coherence for 8+ exchange conversations. Sol's pricing puts Astra-equivalent performance back at the old Turbo cost.
We also have two clients running Opus 5 who couldn't justify Fable 5.1 at $90 output. Opus 5.5 at $45 makes the Fable capability tier affordable again.
The wildcard is Sonnet 5.5 and Haiku 5.5, which Anthropic said are "expected in the coming weeks." If Sonnet 5.5 undercuts Opus 5.5 the way Sonnet 4.7 undercut Opus 4, and if it lands before mid-October, the entire deployment queue reshuffles again.
For now: if you're running Astra or Opus 5 in production and your workload doesn't require the absolute ceiling performance, the new models are worth testing this week. The 40-50% price cut is real, the capability drop is minimal, and the API swap is a two-line config change.