OpenAI cuts GPT-5.6 Luna pricing by 80%

• OpenAI made Luna cheap enough to change routing: GPT-5.6 Luna fell 80% to $0.20 per 1M input tokens and $1.20 per 1M output tokens, while Terra fell 20%, according to OpenAI's announcement.
• Paid Codex and ChatGPT Work budgets did not change, but Luna and Terra now burn fewer credits; thsottiaux's reply says the changes affect both API prices and paid subscriptions.
• Sol's standard price stayed fixed, while API Fast mode offers up to 2.5x speed at 2x price with the same intelligence, per OpenAIDevs' Fast mode note.
• The discount is backed by systems work: OpenAI reported 20% lower serving costs from production GPU kernel improvements and more than 15% better token-generation efficiency in OpenAI's efficiency summary.
• Codex users pushed Luna Max for cheap mechanical work, but pvncher said Luna is not the model to force into Multi agents v2 subagent roles.

The OpenAI pricing post includes the quota catch: subscription prices and quota budgets stay fixed, but Terra and Luna consume fewer credits. The engineering post names the stranger part: Sol in Codex rewrote production kernels in Triton and Gluon, with FpSan checking correctness. Simon Willison's link post says he moved Datasette Agent from Gemini 3.1 Flash-Lite to Luna, while Vercel's changelog says AI Gateway model IDs did not change.

The new API table is simple:

• GPT-5.6 Luna: $0.20 input, $1.20 output per 1M tokens, down 80%, according to sama's price table.
• GPT-5.6 Terra: $2 input, $12 output per 1M tokens, down 20%, according to sama's price table.
• GPT-5.6 Sol Standard: unchanged.
• GPT-5.6 Sol Fast: up to 2.5x Standard speed at 2x Standard price, with no intelligence change, per OpenAIDevs' Fast mode note.
• Existing API requests tagged priority continue to work, according to OpenAIDevs' Fast mode note.

OpenAI framed the cut as efficiency gains passed through to customers, not a new model launch.

The subscription effect is where the release landed for coding-agent users. The official post says ChatGPT and Codex subscription prices and quota budgets remain unchanged, while Terra and Luna consume fewer credits; WesRoth's clarification made the same point for non-API users.

Hangsiin's screenshot put the new Plus plan range at 250 to 2,000 Luna local messages per 5 hours, compared with 10 to 100 Sol messages on the same plan and 200 to 2,000 Sol messages on Pro 20x in the plan table.

OpenAI also reset Codex and ChatGPT Work usage limits for the weekend, with thsottiaux saying the reset was meant to let users run “100,000 Luna threads.”

The first Codex move was not a model selector. It was the reasoning-effort dropdown.

• Max reasoning is off by default in Codex, according to daniel_mac8's PSA.
• Luna is being used for lint fixes, unit tests, and dependency updates, per Hayden Bleasel's task list.
• Hayden Bleasel put his personal split at Luna for deterministic tasks and Sol as the day-to-day workhorse in a follow-up.
• cto_junior's hands-on test found Luna strong at High+ effort for distributed-systems explainers, with code generation sometimes taking more than one attempt to compile.
• simonw's Datasette Agent test said Luna generated SQL, HTML, and JavaScript quickly enough that he switched his public demo instance from Gemini 3.1 Flash-Lite.

Cheap Luna is Christmas morning for coding-agent tinkerers, as long as the job is well specified.

OpenAI also upgraded auto-review in the ChatGPT app and Codex CLI from GPT-5.4 to GPT-5.6 Luna. Combined with the new Luna price, OpenAI expects auto-review to cost about 10x less per OpenAIDevs.

The more interesting phrasing came from thsottiaux, who described “review for me” as Luna being used to prevent high-risk actions from the main agent. That makes Luna the cheap guardrail model inside the workflow, not just the cheap worker model.

The sharpest limitation is orchestration. pvncher said Luna should not be forced into the model catalog for capabilities that are not natively exposed, because Multi agents v2 depends on proactive inter-agent communication that Sol and Terra handle better.

The distinction that emerged in replies:

• Luna is not Multi agents v2 compatible as a Codex subagent, according to pvncher's compatibility reply.
• Thread orchestration works in the app, not the CLI, according to pvncher's CLI reply.
• The current workaround is app thread automation that launches standalone Luna threads and waits on them, per pvncher's workaround reply.
• Threads are pure delegation, while subagents are collaboration, according to pvncher's thread distinction.

Outside Codex direct usage, omarsar0 said Luna works in his orchestrator, where Terra or Luna runs the main executor loop and Sol acts as an on-demand advisor.

OpenAI's engineering post attributes the price cut to work across inference and the agentic harness, not only model distillation.

The concrete levers:

• Load balancing: routing by geography, capacity, accelerator type, context length, cache availability, and request properties, according to OpenAI's engineering post.
• Kernel optimization: GPT-5.6 Sol in Codex rewrote production kernels for the forward pass, with Triton and Gluon named as the GPU programming languages.
• Verification: OpenAI says FpSan, its Floating-Point Sanitizer, helps validate kernels written by GPT-5.6 Sol.
• Speculative decoding: Sol improved its own draft model by designing and running hundreds of architecture experiments, then monitoring training and intervening during hardware failures and instability.
• Harness changes: Codex and ChatGPT Work use a Rust orchestration layer, deferred discovery for tools and plugins, 10,000-token default tool-output caps, append-only model-visible history, and deterministic tool ordering.

OpenAI's summary numbers were 20% lower end-to-end serving costs from kernel work and more than 15% better token-generation efficiency from speculative decoding in the official post.

Artificial Analysis said Luna's Cost per Task fell 80%, Terra's fell 20%, and Terra remained largely behind Luna and Sol on its Intelligence Index versus Cost per Task frontier in its pricing update.

The coding-specific numbers were more uneven:

• DeepSWE: Luna Max matched GPT-5.5 xHigh at 67.2% versus 67.0%, with $0.61 versus $7.23 per task, according to reach_vb's chart.
• Fullstack Code Arena: Sol xHigh ranked #3 in Fullstack and #5 overall, Terra xHigh ranked #10 and #20, and Luna xHigh ranked #14 and #18, per the Arena update.
• Document Arena: Sol ranked #11, Terra #10, and Luna #18, according to Arena's cross-arena follow-up.
• Vision Arena: Sol ranked #14, Terra #25, and Luna #40, according to Arena's cross-arena follow-up.
• Text Arena: Sol ranked #13, Terra #34, and Luna #56, according to Arena's cross-arena follow-up.

The price-performance charts favor Luna hard. The broader arena results show why OpenAI still has Terra and Sol in the lineup.

OpenRouter stacked its own 50% exclusive discount on top of the OpenAI cut, putting Luna at $0.10 input and $0.60 output per 1M tokens and Terra at $1 input and $6 output in OpenRouter's post.

Vercel's AI Gateway changelog says GPT-5.6 Luna and Terra received upstream no-markup pricing, Sol Fast moved from 1.5x to 2.5x faster, and existing model IDs get the new rates with no code change.

Cognition updated FrontierCode for the new discounts, saying the GPT-5.6 series now sits on its price-performance Pareto curve in its leaderboard note. Scott Wu said Devin users should see the cost decreases reflected automatically in his rollout note.

DeepSeek answered the Luna cut with V4 Flash 0731. Artificial Analysis scored it at 50 on the Intelligence Index, one point behind GPT-5.6 Luna Max at 51, while estimating DeepSeek's Cost per Task at roughly 60% lower than Luna because of its 98% cache-hit discount in the DeepSeek evaluation.

The same post lists the model at a 1M-token context window, 284B total parameters, 13B active parameters at inference time, and $0.14 input, $0.28 output per 1M tokens in the model details.

The agent-session caveat is cache-heavy traffic. teortaxesTex's table put Luna cache hits at $0.02 per 1M tokens versus $0.0028 for V4 Flash, and Luna output at $1.20 versus $0.28 for V4 Flash.

One workload report matched that split: rishdotblog saw DeepSeek V4 Flash data-extraction recall rise from 71% to 79%, while Luna Medium was 67% at the same cost after the price cut. In a reply, rishdotblog called DeepSeek better for that job but noted capacity problems with 503 errors.