Glonce

OpenAI cuts GPT-5.6 serving costs 20% with post-launch efficiency work

modelsThe Glonce editorsOpenAI🔥82Glonce rating82Impact82Novelty68Relevance100Surprise62Players95Credibility88

OpenAI released a follow-up write-up describing efficiency work it did on GPT-5.6 after the model had already shipped. The company says it applied GPT-5.6 Sol, the flagship model in the family, to the task of making GPT-5.6 itself cheaper and faster to run in production. The headline results: a 20% reduction in end-to-end serving costs from production GPU kernel improvements, and a token-generation efficiency gain of more than 15% from improved speculative decoding. The kernel work was done in Triton and Gluon, OpenAI's programming languages for writing GPU kernels, and covered routing, scheduling, kernel optimisation and caching across the inference stack; refined load-balancing heuristics also cut serving costs by reducing uneven routing across GPUs. On the agent side, OpenAI describes changes to the agentic harness that GPT-5.6 runs inside: deferred discovery of information the model does not immediately need, a cap of 10,000 tokens on tool output fed back into the model, and prompt caching that treats conversation history as append-only so cached context is reused rather than recomputed. OpenAI frames this as a second axis of the GPT-5.6 release, alongside training-time efficiency: the company says GPT-5.6 was trained to do more useful work per token, and this piece extends that same efficiency goal to the deployed system after launch. The GPT-5.6 family itself ships in three tiers: Sol, the flagship model that also powers OpenAI's Codex product; Terra, a mid-tier model OpenAI says matches the intelligence of the prior GPT-5.5 model at half its price; and Luna, the fastest and cheapest tier, priced 80% below Sol. Listed API pricing puts Sol at $5 per million input tokens and $30 per million output tokens, Terra at $2.50 and $15, and Luna at $1 and $6. The models are rolling out through ChatGPT, Codex and the OpenAI API, with Plus, Pro, Business and Enterprise ChatGPT users able to reach Sol at medium-or-higher reasoning effort settings as the global rollout continues. OpenAI also published benchmark comparisons against Anthropic's Claude Fable 5: on Agents' Last Exam, a professional-workflows benchmark, Sol scores 53.6 against Fable 5's 40.5, a 13.1-point lead, and at medium reasoning effort Sol beats Fable 5 by 11.4 points while using about 25% of the tokens. On the Artificial Analysis Coding Agent Index, Sol at maximum reasoning scores 80, a 2.8-point edge over Fable 5, while using under half the output tokens, finishing in about half the time and at roughly one-third the cost. OpenAI states Terra and Luna outperform Fable 5 at roughly one-sixteenth its estimated cost. On knowledge-work benchmarks, OpenAI reports 92.2% on BrowseComp and 62.6% on OSWorld 2.0, the latter said to surpass Claude Opus 4.8 while using 85% fewer output tokens. On cybersecurity-oriented benchmarks, GPT-5.6 scores 73.5% on ExploitBench 2 versus 47.9% for GPT-5.5; on ExploitGym 3 it reaches a 24.9% pass rate under a two-hour constraint and 33.7% under a six-hour constraint, against 15.1% for GPT-5.5; and on SEC-Bench Pro it scores 71.2% versus 45.8% for GPT-5.5.

Key facts

  • OpenAI cut GPT-5.6's end-to-end serving costs 20% through production GPU kernel improvements written in Triton and Gluon.
  • Improved speculative decoding lifted token-generation efficiency more than 15%.
  • The GPT-5.6 family has three tiers: Sol (flagship, powers Codex), Terra (matches GPT-5.5 intelligence at half its price), and Luna (80% cheaper than Sol).
  • On the Artificial Analysis Coding Agent Index, Sol at maximum reasoning scores 80 versus Claude Fable 5, using under half the output tokens at roughly one-third the cost.
  • Agentic harness changes include deferred discovery, a 10,000-token cap on tool output, and append-only prompt caching.

Why it matters

Most model efficiency work happens before release, during training. Here OpenAI describes a second round of optimisation applied after GPT-5.6 was already deployed, treating cost and speed as targets to keep improving post-launch rather than a fixed property of the shipped model. A 20% cut in serving costs plus a 15%+ gain in generation efficiency, achieved without changing the model's outputs, changes the economics of running GPT-5.6 at scale for anyone already using it.

Who it affects

Developers and enterprises building on the OpenAI API, ChatGPT Plus, Pro, Business and Enterprise users who can reach Sol at medium-or-higher reasoning effort, and users of Codex, which Sol also powers. The pricing tiers, Sol at $5/$30 per million input/output tokens, Terra at $2.50/$15, and Luna at $1/$6, mean the efficiency work indirectly benefits anyone paying for GPT-5.6 access, since lower serving costs sit behind the same sticker prices.

How to use it

GPT-5.6 is available now through ChatGPT, Codex, and the OpenAI API, rolling out globally. Sol is the model to reach for coding and agentic workloads where OpenAI cites its lead on the Artificial Analysis Coding Agent Index; Terra is positioned as a like-for-like swap for GPT-5.5 users at half the price; Luna suits high-volume, cost-sensitive use where speed matters more than top-end reasoning.

How solid is it

The efficiency figures (the 20% and 15%+ numbers) and the benchmark scores against Claude Fable 5, GPT-5.5 and Claude Opus 4.8 are all OpenAI's own reported results; none of the cited benchmarks appear to have independent, third-party reproduction referenced in the source material. The specific engineering detail, naming Triton and Gluon, the 10,000-token tool-output cap, append-only caching, gives the efficiency claims more texture than a bare percentage, which is a point in their favour, but the numbers themselves are self-reported.

Risks and caveats

All comparative benchmark numbers come from OpenAI's own testing of both its models and competitors' models; independent verification is not part of the source material. Cost and speed comparisons against Claude Fable 5 depend on OpenAI's own estimate of Fable 5's serving cost, which is not something OpenAI directly controls or necessarily discloses in full. As with any efficiency-focused update, gains in cost and speed are not the same claim as gains in raw capability, and the write-up is explicit that this is optimisation of an already-shipped model rather than a new capability release.

“After deployment, we applied GPT-5.6 Sol to advance the frontier of efficiency by making itself more efficient to run.”

— OpenAI