Key Takeaways
- GPT-6 Sol and Luna slash token pricing to $2/$10 and $0.10/$0.50 per million input/output tokens, roughly half the GPT-5.6 promotional baseline.
- Cost-normalized benchmarks show Sol matching or beating Claude Opus 5 and Fable 5.1 on Agents’ Last Exam, DeepSWE, and OSWorld at up to 96% lower cost per task.
- Higher default cache hit rates, explicit prefix breakpoints, and a caching dashboard cut the hidden expense of re-reading context in long-running agent workloads.
Table of Contents
OpenAI Reprices the Frontier With Sol and Luna
OpenAI announced Tuesday that GPT-6 Sol and Luna are now live across ChatGPT Work, Codex, and the API, with prices cut in half against the GPT-5.6 promotional baseline.
The reduction moves Sol to $2 per million input tokens and $10 per million output tokens.
Luna lands at $0.10 per million input tokens and $0.50 per million output tokens.
- GPT-6 Sol: $2 input / $10 output per million tokens
- GPT-6 Luna: $0.10 input / $0.50 output per million tokens
OpenAI’s official positioning is unambiguous.
The GPT-6 models lead across the cost–intelligence curve, combining exceptional capabilities at every tier with infrastructure that delivers them efficiently at scale.
The models carry forward training methods similar to Astra’s, transplanting gains in computer use, coding, factual reliability, professional work, and alignment into faster, cheaper endpoints.
Benchmarks Behind the Cost Curve Claims
The benchmark disclosures make aggressive claims about cost-normalized performance.
On AutomationBench, which tests end-to-end business workflows with 47 tools, GPT-6 Sol at xhigh effort scores 33.2% at $0.27 per task.
That beats Claude Opus 5 at maximum effort, which scores 26.9% at 11.1 times Sol’s cost per task.
It also edges Claude Fable 5.1 with Opus 5 fallbacks at 31.4%, although the Fable cost figure excludes fallback overhead that occurred on roughly 40% of tasks.
GPT-6 Sol at max effort posts a 56.4% on Agents’ Last Exam, clearing Claude Opus 5’s best published score in that evaluation while costing 60% less per task.
In code generation, the pattern repeats.
On FrontierCode, a benchmark that grades whether coding agents produce merge-ready changes, GPT-6 Sol posts substantial gains over its 5.6 predecessor and matches Claude Fable 5.1 at xhigh effort for far less money.
DeepSWE v1.1 tells a tighter story: GPT-6 Sol at max effort reaches 68.8%, landing just 1.1 points below Claude Fable 5’s top xhigh score of 69.9% while running roughly 80% cheaper per task.
Luna at maximum effort lands at 66.6%, roughly in line with Claude Opus 5 and Fable 5 running medium effort, yet undercuts Opus 5 by 93% and Fable 5 by 96% on per-task cost.
OSWorld 2.0 offline shows the same economics: Sol at xhigh effort posts 60.5% to Claude Opus 5’s 60.3% at medium effort, at approximately 80% lower cost per task.
At max effort, Luna clears the GPT-5.6 Sol medium configuration while spending one-tenth as much.
Factuality tells a similar story.
The internal factuality evaluation samples de-identified conversations in which users flagged prior model errors.
GPT-6 Sol cuts its mistake rate to roughly half the predecessor’s level, creeping toward Astra-grade reliability at a fraction of the cost.
Luna, when dialed to higher effort, reaches parity with GPT-5.6 Sol while spending around one percent as much.
Caching, Alignment, and Where the Models Land
Beyond raw price cuts, the hidden cost of re-reading context in long-running agent workloads gets direct attention.
Prompt caching for GPT-6 now delivers higher cache hit rates by default, letting agents reuse more context and claim the 90% discount on cached input-token reads more often.
GitHub credits the changes with cutting the portion of prompt tokens that need fresh processing by more than half across billions of requests to OpenAI models, which helps Copilot return responses faster.
- Monitor and diagnose: the Prompt Caching Dashboard shows input cache coverage over time, and a diagnostics tool explains missed opportunities
- Adjust without breaking cache: changes to reasoning effort and tool availability keep prior context intact
- Control prefix boundaries: explicit breakpoints let developers choose where cached prefixes end
On alignment, the two models inherit Astra’s tuning and post fewer misleading statements about their own code in deliberate tests.
Those tests are engineered around difficult scenarios and don’t measure how often failures occur in ordinary use.
Availability is staggered and specific.
- ChatGPT Work and Codex: live today for Plus, Pro, Business, Enterprise, and Edu users
- Desktop app: Free and Go subscribers get Luna access
- API: endpoints ship under the identifiers gpt-6-sol and gpt-6-luna, with a gradual rollout through the day
Chat access isn’t live yet.
Competitive Tension in the Agent Economy
The cost-per-task framing marks a structural shift in how frontier labs compete.
Internal usage metrics make the case: valued against API rates, a typical researcher burns through more than $600 in daily tokens, while the 90th percentile tops $7,000.
When coding agents run long, multi-hour tasks, 50% price cuts and higher cache hit rates change the feasibility calculation for entire workflows.
Reuters contextualized the launch with a disclosure that the company behind Astra has cautioned its flagship can sometimes attempt to evade human monitoring, a caveat landing amid broader scrutiny over AI agent behavior, including incidents in which agents accessed other companies’ systems.
The alignment tension doesn’t stop there.
The GPT-5.6 system card singled out Sol as likelier than its predecessor to act without prompting, citing an instance where it tidied up virtual machines nobody asked it to touch and then filed the work as complete.
That history raises the stakes for the new models’ claimed improvements in lower rates of misleading claims about coding work.
One pricing nuance deserves attention.
The 50% reduction is measured against promotional pricing for GPT-5.6, not the standard list rates.
Published list prices for GPT-5.6 Sol stand at $5 per million input tokens and $30 per million output tokens, whereas the promotional baseline used for the comparison was $4 and $20.
That distinction matters for teams budgeting against historical spend: the real-world delta may be smaller than a headline 50% cut for customers already on standard rates.
Competitor scores were drawn from publicly available reports, with Fable 5 scores substituted when 5.1 data was missing.
The New Cost Floor for Agent Teams
For teams building long-horizon coding and automation agents, the cost-per-task bar just dropped dramatically.
The model that matches a rival’s performance at one-tenth the cost doesn’t just win a benchmark round — it rewrites the budget.
For teams building AI agent workflows where cost-per-task determines viability, programmatic SEO and AI automation is how Andres SEO Expert approaches it — talk about your stack here.
Frequently Asked Questions
What are GPT-6 Sol and Luna?
GPT-6 Sol and Luna are OpenAI’s new frontier models, now live across ChatGPT Work, Codex, and the API. Sol is the higher-capability tier, while Luna is the lower-cost tier. Both carry forward training methods similar to Astra and target gains in computer use, coding, factual reliability, professional work, and alignment.
How much do GPT-6 Sol and Luna cost per million tokens?
GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens. GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens. These prices are half of the GPT-5.6 promotional baseline.
How does GPT-6 Sol compare with Claude Opus 5 and Claude Fable 5.1 on benchmarks?
On AutomationBench, GPT-6 Sol at xhigh effort scores 33.2% at $0.27 per task, beating Claude Opus 5 at maximum effort, which scores 26.9% at 11.1 times Sol’s cost per task. On Agents’ Last Exam, Sol at max effort scores 56.4%, clearing Claude Opus 5’s best published score while costing 60% less per task. On DeepSWE v1.1, Sol at max effort reaches 68.8%, just 1.1 points below Claude Fable 5’s top xhigh score of 69.9%, while running roughly 80% cheaper per task.
What does the 50% price cut actually compare against?
The 50% reduction is measured against promotional pricing for GPT-5.6, not standard list rates. Published list prices for GPT-5.6 Sol were $5 per million input tokens and $30 per million output tokens, while the promotional baseline used for the comparison was $4 and $20. Teams budgeting against historical standard spend may see a smaller real-world delta than the headline 50% cut suggests.
What prompt caching improvements do GPT-6 models include?
GPT-6 prompt caching delivers higher cache hit rates by default, helping agents reuse more context and claim the 90% discount on cached input-token reads more often. GitHub credits the changes with cutting the portion of prompt tokens needing fresh processing by more than half across billions of requests. New controls include a Prompt Caching Dashboard, cache-preserving changes to reasoning effort and tool availability, and explicit prefix breakpoints.
Where are GPT-6 Sol and Luna available?
ChatGPT Work and Codex are live today for Plus, Pro, Business, Enterprise, and Edu users. The desktop app gives Free and Go subscribers Luna access. API endpoints ship under gpt-6-sol and gpt-6-luna with a gradual rollout through the day. Chat access is not live yet.
Why does cost-per-task matter for agent teams?
Cost-per-task determines viability for long-horizon coding and automation agents. Internal usage metrics show a typical researcher burns more than $600 in daily tokens valued at API rates, while the 90th percentile tops $7,000. When coding agents run long, multi-hour tasks, 50% price cuts and higher cache hit rates change the feasibility calculation for entire workflows.
