GPT-6 Sol and Luna cut prices in half, but benchmarks against rivals come with caveats

OpenAI expanded the GPT-6 family with two models focused on cost-effectiveness: GPT-6 Sol and GPT-6 Luna. They arrive a few weeks after the launch of GPT-6 Astra, which the company keeps as the most capable model in the line. Sol and Luna don't try to outperform Astra in raw intelligence: the stated goal is to bring part of Astra's architecture and alignment advances to cheaper tiers of the catalog, tackling what OpenAI calls the "cost-intelligence curve."

The announcement came with a battery of benchmarks comparing Sol and Luna against Claude Opus 5 and Claude Fable 5.1, from Anthropic. For those deciding which model to put behind a coding agent or automation in production, the question that matters is simple: do the numbers hold up the hype, or is OpenAI bending the ruler?

Price dropped, but dropped from where

The price cut is the most concrete data point in the announcement, and it's documented with precision:

Values per 1 million tokens.

It's worth noting a detail: the comparison is against the "promotional price" of the GPT-5.6 generation, not against the original list price. This doesn't invalidate the cut, but it means the real percentage gain since the initial launch of 5.6 may differ from what the table suggests at first glance.

The benchmarks OpenAI chose to show

On AutomationBench (maintained by Zapier, which tests agents in real workflows with 47 tools spanning sales, marketing, operations, support, finance, and HR), GPT-6 Sol in xhigh mode scored 33.2% accuracy at a cost of $0.27 per task. For comparison, Claude Opus 5 in max mode came in at 26.9%, costing 11.1 times more per task. Claude Fable 5.1 (with fallback to Opus 5) reached 31.4%, but OpenAI itself admits, in a footnote, that the cost of the fallback to Opus 5 is not accounted for, and that this fallback occurred in about 40% of tasks. In other words: the cost figure for the strongest direct competitor in this test is understated by construction, which artificially inflates Sol's cost advantage on this specific line.

On Agents' Last Exam, which evaluates complex professional workflows, Sol in max effort reached 56.4%, above Opus 5's best mark on the same test, at 60% less cost per task.

In coding, DeepSWE v1.1 (software engineering tasks on real repositories) put Sol at max at 68.8%, 1.1 percentage points from Claude Fable 5's best mark (69.9% at xhigh), but at roughly 80% less cost per task. Luna, on the same test, reached 66.6% at max, comparable to Opus 5 and Fable 5 at medium effort, costing 93% less than Opus 5 and 96% less than Fable 5.

On FrontierCode, which measures whether a code agent's changes are ready to merge into real codebases, Sol improved substantially over GPT-5.6 Sol and tied technically with Fable 5.1 at xhigh, at a much lower cost. In computer use, on OSWorld 2.0 (maintained by xlang.ai), Sol at xhigh scored 60.5% against Opus 5's 60.3% at medium effort, at roughly 80% less cost per task.

The fine print that changes the reading

The pattern that repeats across almost every chart is the same: OpenAI compares Sol/Luna's best effort against a lower or medium effort from Anthropic's models, or vice versa, depending on which comparison favors the result. This is a common practice in vendor benchmark reports, but the practical effect is that "60% cheaper for the same result" doesn't always mean the same effort configuration on the other side, and OpenAI itself acknowledges this: "scores for competitor models were taken from publicly available reports," meaning Anthropic's numbers weren't reproduced by OpenAI in a controlled environment, but instead taken from the competitor's own public disclosures.

This doesn't make the numbers false, but it changes the weight an engineering team should give them before switching models in production. The practical recommendation is to run your own internal benchmark with the team's real workload before migrating, rather than deciding based solely on the announcement's table.

A second point of attention: the response-style comparison (the prompt example about redesigning a website in React) is subjective by definition, and OpenAI itself states its preference ("we prefer GPT-6 Sol's reply here"). It's an interesting qualitative data point, not a measurement.

Cheaper cache for long-running agents

The part of the announcement with the most direct impact on operating cost for those already running agents in production is prompt caching. OpenAI says it has improved the default cache hit rate, with a 90% discount on input tokens read from cache. Two new features also arrived: adjusting the reasoning level or enabling/disabling tools mid-conversation without breaking the cache, and explicit breakpoints to control where the cached prefix ends.

GitHub, cited in the announcement, reports that these improvements reduced by more than 50% the fraction of prompt tokens that require processing from scratch, across billions of requests to OpenAI's models, helping Copilot respond faster. This is third-party data (GitHub) that lends some independent backing to the announcement, though without detailing the measurement methodology.

What changes in practice for builders

For teams already using OpenAI's API in code agents (Codex) or automation, the 50% price cut on Sol and Luna, combined with the cache discount, opens real room to run more iterations on the same budget, especially in high-volume, low-complexity tasks, where Luna (cheaper, at $0.10/$0.50 per million tokens) already delivers results comparable to the previous GPT-5.6 Sol at a fraction of the cost, according to OpenAI's internal factuality evaluation.

The announcement itself includes a revealing data point about internal consumption: daily token usage in code agents at OpenAI would already cost, at API pricing, more than $600/day for the median researcher and more than $7,000/day for the 90th percentile. This is useful context for calibrating expectations: long-running code agents consume tokens at a scale that makes the price per million tokens an architecture decision factor, not a billing detail.

GPT-6 Astra remains OpenAI's own recommended option for the most demanding cases, and it stays outside the scope of this announcement's price cut. Sol and Luna are, in practice, the company's bet to dominate the price range where Anthropic has positioned itself with Claude Opus 5 and Claude Fable 5.1.

Availability

GPT-6 Sol and Luna have been available since the announcement in ChatGPT Work and in Codex for Plus, Pro, Business, Enterprise, and Edu users. Users on the Free and Go plans have access to Luna through the desktop app. The models haven't reached Chat yet. In the API, they're available under the identifiers gpt-6-sol and gpt-6-luna. OpenAI notes that the rollout in ChatGPT is gradual throughout the day, so the models' absence from the interface may be temporary.

Is the hype worth it?

The price cut is real and verifiable. The cache efficiency gain also has backing in third-party data (GitHub). But the narrative of "outperforms Claude Opus 5 at a fraction of the cost" needs to be read with the same skepticism that would apply to any benchmark published by the vendor itself: the effort levels compared vary from chart to chart, the competitor's numbers come from public reports rather than direct testing, and at least one of the comparisons most favorable to Sol (AutomationBench against Fable 5.1) omits a fallback cost that occurred in nearly half of the tasks. For architecture decisions, the safest path remains measuring the team's real workload against competitors before migrating, rather than blindly trusting the launch table.