The Frontier Is a Pricing Tier Now

**Subtitle:** GPT-6 Sol and Luna match yesterday's flagships at a fraction of the cost — and the fine print is the interesting part.

Hacker News picked it up fast: OpenAI’s “GPT-6 Sol and Luna” announcement hit the front page with 1,704 points and 814 comments in a day — the biggest AI story of the week. The headline claim is simple: smaller, cheaper models, 50% off. The interesting part is what the numbers quietly confess.

What was announced

Earlier this month OpenAI launched GPT-6 Astra, its flagship. Sol and Luna are the mid-tier follow-ups: GPT-6 Sol at $2/$10 per million input/output tokens, GPT-6 Luna at $0.10/$0.50. That’s 50% cheaper than the GPT-5.6 promotional pricing they replace. Both are available in ChatGPT Work and Codex starting today, plus the API as gpt-6-sol and gpt-6-luna. (All figures below come from OpenAI’s announcement post, unless noted.)

None of this is surprising — every lab now ships a three-tier lineup. What’s worth reading is the benchmark appendix, because OpenAI published cost-per-task numbers alongside the scores, something vendors rarely volunteer.

The numbers, honestly presented

OpenAI evaluated its own models in its research environment; competitor scores were taken from public reports. Read with that in mind:

• AutomationBench (end-to-end business workflows across 47 tools): GPT-6 Sol at xhigh effort scores 33.2% at $0.27 per task. Claude Opus 5 at max effort scores 26.9% at 11.1x that cost.
• DeepSWE v1.1 (long-horizon software engineering in real codebases): GPT-6 Sol hits 68.8%, within 1.1 points of Claude Fable 5’s best 69.9%, at roughly 80% lower cost per task.
• GPT-6 Luna at max effort scores 66.6% on DeepSWE — comparable to Claude Opus 5 and Fable 5 at medium effort — at 93% and 96% less cost per task respectively.
• On OpenAI’s internal factuality eval, Luna at higher effort matches GPT-5.6 Sol at “about a hundredth its cost.”

Then the footnotes. OpenAI admits the Claude Fable 5.1 cost figure understates reality because it omits the cost of Opus 5 fallbacks — which happened on ~40% of tasks. The alignment numbers come from OpenAI’s own internal evaluations of deliberately adversarial situations. None of this is fraud; it’s marketing-shaped measurement. The direction is real, the exact ratios are not load-bearing.

The real story: the collapse

Strip the marketing and one pattern survives every vendor’s cherry-picking: last generation’s flagship is this generation’s mid-tier. A model that costs $0.10 per million input tokens now matches, on several agentic benchmarks, models that cost two orders of magnitude more just months ago.

This is the honest version of “democratization.” The frontier doesn’t disappear; it becomes a pricing tier. The practical question for anyone building is no longer “which flagship?” but “how much of my workload actually needs the flagship?”

Agents are token furnaces — and the 50% that matters more

The most honest number in the whole announcement isn’t a price at all. Buried in the coding section: valued at API prices, OpenAI’s own researchers’ daily token usage exceeds $600 for the median researcher and $7,000 for researchers at the 90th percentile. That is what sustained agentic coding actually costs — at the company that builds the models.

Against that, the headline 50% price cut is the smaller story. The bigger one is caching: prompt-caching improvements now deliver 90% discounts on cached input-token reads, and GitHub reports that the share of prompt tokens requiring fresh processing fell by more than 50% across billions of requests. For long-running agents that re-read the same context all day, cache hit rate will move your bill more than any price list ever will.

(Also note: the 50% cut is measured against GPT-5.6’s promotional pricing. The baseline was already discounted, so the real cut is softer than the headline.)

What to do with this

1. Default to the mid-tier. Route everything to Sol/Luna-class models and escalate only the tasks that demonstrably fail. The benchmark gap between flagship and mid-tier is now single-digit percentage points on several coding benchmarks; the cost gap is 10x.

2. Measure on your own workload. Vendor benchmarks are adversarial marketing. Run your own eval set before you believe a cost-per-task table — especially one where the competitor’s fallback costs are footnoted away.

3. Optimize caching before shopping prices. Explicit cache breakpoints and effort controls that preserve cache reuse are the difference between a $600/day habit and a $60 one.

4. Expect this to keep happening. Inference costs are falling faster than sticker prices, which means today’s “generous” 50% cut is tomorrow’s baseline. Never sign a long-term contract anchored to today’s per-token price.

Bottom line

GPT-6 Sol and Luna are genuinely good news for builders: more intelligence per dollar, with the receipts to prove the direction. Just don’t mistake a pricing announcement for a capability announcement. The models got cheaper because the infrastructure got cheaper — and OpenAI kept the difference it didn’t pass on. That’s not a complaint; it’s the business. Your job is to capture as much of the falling cost curve as you can, with your own measurements, before the next tier shuffle moves the goalposts again.