OpenAI has rolled out GPT-6 Sol and a smaller companion, GPT-6 Luna, widening the GPT-6 family just weeks after the flagship GPT-6 Astra arrived, and the company is making the economics of running advanced AI a central selling point rather than an afterthought. Both new models are priced 50% below the promotional rates of their GPT-5.6 predecessors, and OpenAI is framing the move as a direct response to pressure from Anthropic’s recently released Claude Opus 5.5 and Claude Fable 5. With enterprises increasingly wiring agents into long-running workflows, the argument is that price-performance, not leaderboard position, is increasingly what wins contracts.
GPT-6 Sol Cuts Token Pricing in Half
The headline number for GPT-6 Sol is $2 per million input tokens and $10 per million output tokens, down from $4 and $20 on GPT-5.6 Sol. That 50% cut applies to both directions of the pipeline, which matters for agent workloads that produce large amounts of generated text, code or tool calls. Luna goes considerably further, with input tokens priced at $0.10 per million and output tokens at $0.50 per million, roughly matching the cheaper end of the frontier market. Together, the two models give OpenAI a clearer three-step ladder: Astra at the top for maximum capability, GPT-6 Sol in the middle for demanding coding and agentic work, and Luna for high-volume applications where cost dominates.
GPT-6 Sol: Agent Economics Become Harder to Ignore
Token pricing only describes part of the bill. Agents carry context from one step to the next, repeatedly processing the same instructions, tool descriptions and intermediate outputs, which inflates both latency and cost. OpenAI says it has overhauled prompt caching for the GPT-6 line, claiming default cache hit rates that let agents reuse more context and unlock discounts of around 90% on cached input reads. New developer tools expose cache diagnostics and let teams control which prompt prefixes get cached, turning caching from an opaque backend feature into a tunable cost lever. OpenAI added that caching improvements over recent months have already cut fresh-processing share of prompt tokens by more than 50% across billions of requests, helping GitHub Copilot respond faster.
Coding Benchmarks Bring Cost Into Focus
Coding agents now chew through entire repositories rather than spitting out snippets, so token bills scale with task length. OpenAI disclosed that daily internal usage has exceeded $600 for the median researcher and $7,000 for those at the 90th percentile. On DeepSWE v1.1, a long-horizon software engineering benchmark, GPT-6 Sol at maximum effort scored 68.8% versus Claude Fable 5’s 69.9% at xhigh effort, a gap of 1.1 percentage points. OpenAI says GPT-6 Sol delivered that result at roughly 80% lower cost per task, and that Luna, at 66.6%, ran 93% cheaper than Claude Opus 5 and 96% cheaper than Claude Fable 5 in matched configurations.
GPT-6 Sol Pressures Anthropic’s Premium Tier
The same pattern shows up in OpenAI’s AutomationBench results across sales, marketing, operations, support, finance and HR. GPT-6 Sol at xhigh effort scored 33.2% at about $0.27 per task, while Claude Opus 5 at maximum effort scored 26.8% and cost 11.1 times more. OpenAI also reports that Sol beat Claude Fable 5.1 with Opus 5 fallback on the same benchmark at substantially lower cost. The caveats are real: OpenAI ran its own evaluations while competitor numbers came from public reports, and production ChatGPT traffic behaves differently from API calls. Still, the framing matters, because the conversation is shifting from raw capability toward capability per dollar.
What the New Pricing Means for Enterprise Buyers
For CIOs and platform teams, the practical question is whether a slightly higher benchmark score is worth a several-fold increase in operating cost, and whether the cheapest model can reliably finish the job. OpenAI’s 1.05-million-token context window and 128,000-token output ceiling on Sol and Luna make both relevant for large-document and long-session tasks, while the price cuts lower the barrier to running them continuously. The deeper signal is competitive: with Anthropic’s Opus 5.5 and Fable 5 raising the capability bar, GPT-6 Sol is OpenAI’s answer that price-performance, not raw leaderboard position, is now the main battleground in the enterprise AI race.
Enterprise procurement teams are reacting to the new GPT-6 Sol pricing with cautious optimism, noting that the sub-$2 per million input token rate brings OpenAI into closer parity with Anthropic’s recently discounted Claude Opus 5.5 tier and undercuts Google’s Gemini 3 Pro enterprise SKU by roughly 18%. Channel partners report a surge in RFP activity from Fortune 500 buyers seeking to lock in multi-year commitments tied to API price floors, while several mid-market SaaS vendors have already announced plans to migrate their retrieval and summarization pipelines onto Sol to capture margin improvements of 8–12%. On the supply side, OpenAI’s expanded deal with CoreWeave for an additional 1.2 GW of H-class GPU capacity, paired with the long-rumored Oracle cloud backstop, suggests the company is willing to absorb thin gross margins to defend installed base against Fable 5’s open-weight variants, which are gaining traction in regulated industries where on-prem inference is non-negotiable. The broader market read is that GPT-6 Sol is less about chasing a benchmark crown and more about resetting the unit-economics conversation for generative AI deployments at scale.