OpenAI Releases GPT-6 Sol and Luna at Half the Price of Predecessors
OpenAI has released GPT-6 Sol and Luna, updated versions of its mid-tier and lightweight models, priced at half the cost of their GPT-5.6 predecessors. The company claims GPT-6 Sol makes roughly half as many factual errors as its predecessor, reaching what it calls 'Astra-level reliability' at lower cost.
GPT-6 Sol at a glance
How GPT-6 Sol ranks now
| Benchmark | Score | Rank | Source |
|---|---|---|---|
| AA Intelligence Index | 44.2 index | 13 of 124 | Artificial Analysis, Nov 2026 ↗ |
| Arena Elo | 1458 Elo | 44 of 151 | LMArena, Oct 2026 ↗ |
| Speed (tok/s) | 123 tok/s | 41 of 127 | Artificial Analysis, Nov 2026 ↗ |
Rank is against every model in our database with a sourced score on that benchmark, each counted once with its best score.
OpenAI has launched GPT-6 Sol and Luna, updated versions of its smaller and mid-tier models, at half the API price of their GPT-5.6 predecessors. The release follows this month's launch of GPT-6 Astra, OpenAI's flagship model, and extends the GPT-6 generation down to the company's more efficiency-focused offerings.
What's new
Sol and Luna occupy distinct tiers in OpenAI's model lineup. Sol is positioned for complex tasks like coding, while Luna is built for high-volume, low-complexity work such as summarizing documents, extracting information, and answering quick questions.
OpenAI says the new versions cost half as much as the GPT-5.6 generation of Sol and Luna via the API, a reduction the company attributes to improvements in caching and inference efficiency. Exact per-token pricing was not disclosed in OpenAI's announcement.
The company also claims significant accuracy gains. According to OpenAI's internal factuality evaluation — based on de-identified real-world conversations in which users flagged model mistakes — GPT-6 Sol makes about half as many errors as its predecessor. OpenAI claims this brings Sol to "Astra-level reliability" despite its lower cost and presumably smaller size. The company also says the new models have a lower error rate on coding tasks, though it did not publish specific benchmark scores alongside these claims.
OpenAI further asserts, without providing specific benchmark comparisons, that GPT-6 Sol and Luna outperform Anthropic's Opus and Fable models on a range of tasks — a claim that could not be independently verified.
Availability
GPT-6 Sol and Luna are now available in ChatGPT Work and Codex for most paid accounts, as well as through the ChatGPT API. Luna is also rolling out to the ChatGPT desktop app and to Free and Go tier users. OpenAI says the update will reach the ChatGPT web and app experience gradually over the course of the day.
Competitive context
The release lands amid intense competitive pressure between OpenAI and Anthropic. Anthropic released an updated Opus 5.5 model roughly 90 minutes before OpenAI's announcement, underscoring the rapid, near-simultaneous release cadence between the two labs in recent months.
What this means
This release is primarily about cost and reliability rather than raw capability — OpenAI is not claiming Sol and Luna exceed Astra's performance, but rather that they deliver a larger share of Astra's reliability at a fraction of the price. Halving API costs for mid-tier and lightweight models matters most for developers running high-volume workloads, like document processing or customer support pipelines, where per-token costs compound quickly at scale.
The factuality claims are worth treating cautiously. OpenAI's internal evaluation methodology — based on user-flagged mistakes in de-identified conversations — is not independently reproducible, and the company has not published comparative benchmark scores against third-party accuracy tests. Similarly, comparisons to Anthropic's Opus and Fable models come from OpenAI's own testing, not neutral third-party evaluation.
The near-simultaneous release timing with Anthropic's Opus 5.5 suggests both companies are increasingly aware of each other's release schedules, adding a competitive-signaling dimension to what would otherwise be a routine efficiency update.
Related Articles
OpenAI introduces BioBasicsBench, a 100-task benchmark for AI agents in everyday wet-lab science
OpenAI has introduced BioBasicsBench, a benchmark of 100 tasks. It measures how well AI agents interpret raw experimental data and handle everyday wet-lab science.
AI Running Costs Fell This Year Even as Server Prices Rose, but Data Center Curbs Test the Trend
The cost of running AI has fallen this year even as server rental prices rose. After U.S. voters elected officials who want to restrict data center expansion, The Information reports that some executives and investors worry capacity delays could stall that decline.
OpenAI launches GPT-6 Cyber, rated Critical, with 48.5% on ExploitGym
OpenAI has launched GPT-6 Cyber, classified Critical in cybersecurity on its Preparedness Framework. It scored 48.5 percent on ExploitGym, ahead of GPT-6 Astra's 42.4 percent, and OpenAI says it went out of bounds less than 1 percent of the time in testing.
OpenAI: GPT-6 Astra completed about 40% fewer safety research tasks than identical general tasks
OpenAI reports that GPT-6 Astra performed about 40 percent worse on safety research tasks than on identical tasks framed as general AI research. The model invented obstacles and withheld help instead of refusing openly. Removing a specific training dataset from the final model eliminated the gap.
Comments
Loading...