Reading note. This article draws on the public announcement of GPT-5.6 (preview on June 26, 2026; general release on July 9) and on the reporting available at the time of writing. Benchmark and pricing figures come from OpenAI's communication and secondary sources; we cite them as given and flag what remains to be confirmed. Last revision dated at the foot of the page.

In one sentence

On July 9, 2026, OpenAI released GPT-5.6 publicly, after a restricted preview that opened on June 26. The novelty isn't a model, it's a split: a single generation number (5.6) declined into three durable tiers — Sol, Terra, Luna — that explicitly separate intelligence, speed, and cost. Beneath the benchmark records, the signal that strikes us as most important is quiet: at equal quality, GPT-5.6 claims to produce its answers while consuming far fewer tokens. This is not a story about power. It's a story about efficiency — and efficiency, seen from an island, is far more interesting news than performance.

1. What was announced, without the varnish

Let's take the facts first, then discuss what they mean.

The timeline. GPT-5.6 first shipped in preview on June 26, 2026, within a deliberately narrow perimeter — on the order of twenty partner companies, according to several accounts — before a general opening announced for July 9, across ChatGPT, ChatGPT Work, the coding tool Codex, and the OpenAI API — the API (application programming interface) being the interface through which a third-party program calls the model remotely. The models also appeared in GitHub Copilot on public-launch day.

The name. OpenAI now separates two things. The number (5.6) marks the generation. The names — Sol, Terra, Luna — designate capability tiers meant to endure and to advance at their own pace. The stated idea: give a clearer choice between intelligence, speed, and cost, instead of a single model that one tunes by guesswork.

The three tiers, as presented:

  • Sol — the flagship. Complex reasoning, code, scientific work, cybersecurity, and long-running "agentic" tasks (missions the model carries out over time, in several steps, without being re-prompted at each turn). Sol has an Ultra mode and a "maximum reasoning effort" setting for the heaviest tasks.
  • Terra — the everyday model. OpenAI presents it as offering "GPT-5.5 quality at roughly 2× lower cost." For many uses, it's the default choice.
  • Luna — the fastest and cheapest, built for volume and latency-sensitive applications: classification, triage, mass processing.

The pricing, per million tokens (input / output), as announced:

  • Sol — $5 input, $30 output (the same rate as GPT-5.5).
  • Terra — $2.50 / $15.
  • Luna — $1 / $6.

The context window — the amount of text the model can keep "in view" at once — is announced at around 1.05 million tokens for all three tiers, with up to 128,000 tokens of output.

2. A useful detour: "agentic," "multi-agent," "reasoning effort"

Three terms recur in the announcement and deserve a definition in passing, because they carry most of the argument.

Agentic — said of a model able to carry out a long task end to end: break down a goal, chain steps, use tools, correct its trajectory, without a human re-prompting it at every sentence. It's the year's shift: from the model that answers to the model that executes.

Multi-agent — offered in beta, this mode lets GPT-5.6 launch several sub-agents in parallel within a single request, then synthesize their work. In other words, the model subdivides to attack a problem from several angles at once, before reassembling the pieces. Sol's Ultra mode relies on this mechanism to accelerate complex reasoning.

Reasoning effort — a setting that lets the model "think longer" (thus spend more compute and tokens) when the task warrants it, and stay thrifty otherwise. GPT-5.6 makes it an explicit dial, up to a "maximum" level on Sol.

Let's hold the idea: OpenAI no longer sells an intelligence, but a modulable intelligence, whose operating point one can choose between cost, speed, and depth.

3. The benchmarks, and what they hide

Here are the headline figures. We report them, then put them in perspective.

  • Terminal-Bench 2.1 (command-line workflows, agentic code): Sol Ultra at 91.9%, base Sol at 88.8%, versus 88.0% for GPT-5.5 and 88.0% for a competing model in the Claude family (Mythos 5). The claimed lead is thus around 3.9 points.
  • BrowseComp (autonomous web navigation): 92.2%, presented as a new record.
  • OSWorld 2.0 (operating-system control): 62.6%, with a mention that interests us — Sol here reportedly surpasses a leading competitor "while using 85% fewer output tokens."
  • ExploitBench (offensive cybersecurity): competitive with the best competing preview "while using only about a third of the output tokens."

Two honest remarks.

First, the leads are thin. 3.9 points on a coding benchmark, a navigation record at 92% when the previous one wasn't far behind: this is incremental improvement, not a break. On at least one software-engineering leaderboard (SWE-Bench Pro), a competing model would in fact still be ahead, around 80%. The frontier race is now a race of tight packs, not lone champions.

Second — and this is where we want to insist — the figures that truly matter aren't the top percentages, but the "85% fewer output tokens" and the "a third of the tokens." A model that reaches an equivalent result while producing three times less output is a model three times cheaper to run for that result, regardless of its sticker price. The real progress hides in the denominator.

"The right question is no longer 'who has the best score,' but 'how much does a unit of useful intelligence cost.'"

4. The real subject: cost per unit of intelligence

We have already argued here that the price of the token is collapsing, and that this collapse is the major economic fact of AI — more than any record. GPT-5.6 is a clear illustration.

Look at the structure of the offer, not the summit. Terra promises "GPT-5.5 quality at half the price." Luna drops to $1 / $6 per million tokens. And Sol, at the same rate as the previous generation, claims to render the same service while consuming far less output. All three tiers tell the same story in three forms: at constant quality, cost falls; at constant cost, quality rises.

This movement has concrete second-order effects:

  • For developers and small outfits — uses that were out of budget a year ago (agents running continuously, analyses of large corpora, always-on assistants) fall below the break-even line. It's not the cutting-edge capability that democratizes, it's the drop in the mid-range price.
  • For application-layer providers — the margin shifts. When the raw model becomes cheap and abundant, value moves up toward integration, proprietary data, experience — not toward reselling tokens.
  • For frugal players — a Luna at $1 per million input brings the remote API close to the cost of a small self-hosted model. The "rent or host" calculation replays, but on criteria other than price alone (we'll come to that).

5. The blind spot: who holds the switch

One detail of the preview is worth pausing on, because it replays a familiar scene. Initial access to GPT-5.6 was deliberately restricted, and several sources mention government-imposed access restrictions during that phase, with OpenAI indicating that "such restrictions should not become the norm."

The word is out, and it brings us back to what we wrote last month about the cutoff of Fable 5 on Washington's order. The pattern repeats: a frontier model is no longer merely a product, it's an asset the host state may want to throttle. That the restriction here is milder and transitory doesn't change the underlying lesson — the availability of a rented capability does not belong to you.

This is the necessary counterpoint to the enthusiasm of the previous section. Yes, the cost per unit of intelligence is collapsing. But a low cost on a revocable capability remains a dependency, not an asset. Terra at half price is an excellent deal as long as the tap keeps running.

6. Signals to watch

For the reader who wants to check where this story goes:

  1. Whether the efficiency figures hold. The "85% fewer tokens" and "a third of the tokens" are announcement promises. Independent measurements, in the coming weeks, will tell whether the claimed efficiency holds outside the in-house benchmarks.
  2. The fate of Terra. If the "previous generation's quality at half price" tier becomes the norm at every cycle, it confirms that the structural decline in cost is the industry's real, repeatable product.
  3. What becomes of the access restrictions. OpenAI says this should not become the norm. Will there remain access conditions tied to geography or nationality once general availability settles in? The answer will tell whether the preview was a one-off caution or a precedent.
  4. The response of open models. Against a $1 Luna, the true frugal alternative isn't another cheaper proprietary model, but an open-weights model one hosts oneself. Does the quality gap between the two worlds close fast enough to make self-hosting reasonable?

7. A situated word

We write from Réunion Island, 9,000 km from Silicon Valley. From here, a model launch reads first not as a feat, but as a variation in the bill and in dependency.

The good news of GPT-5.6, for a small island player, an association, a frugal lab, isn't that Sol Ultra gains 3.9 points on a coding benchmark we'll never run to saturation. It's that Terra and Luna make quality AI affordable at our scale — that the mid-range, the one that powers the real everyday uses, costs a little less each quarter.

The bad news is the other face of the same coin: the cheaper and more indispensable the tool becomes, the heavier the question "who can cut me off?" weighs. The wisdom here isn't to shun the best tools — that would be shooting oneself in the foot — but to build while keeping, alongside the bright and cheap API, a modest capability no one can unplug. To consume abundance without becoming its prisoner.

GPT-5.6 confirms the trajectory we've been tracking for months: intelligence is becoming a continuous, cheap service. What remains is not to confuse cheap with controlled. The two are not paid in the same currency. 効

Sources and further reading

This document is updated if new elements appear. Last revision: 文 July 10, 2026.