The new GPT-5.6 family: Luna, Terra, Sol

Simon Willison · email · 2026-07-10 Reader 🖥️

AI Unit Economics Evals Frontier Model Convergence / Commoditization Frontier-Lab Hype Long-Running & Managed Agents MCP (Model Context Protocol)Multi-Agent Orchestration Token Efficiency

OpenAI's new GPT-5.6 family (Luna/Terra/Sol) leads on long-running agentic benchmarks and is pitched on token efficiency, but its benchmark story is mixed and partly self-serving and in hands-on coding it is merely competitive with Claude Fable 5 — so the more consequential news is the new API primitives (programmatic tool calling, native multi-agent, explicit prompt caching) that adopt patterns Anthropic pioneered.

practical · mixed — A practitioner's launch-day appraisal (expository reporting of specs/pricing/benchmarks fused with a lightly argumentative adopter's verdict). Extracts only the GPT-5.6 family section of the newsletter; the sqlite-utils 4.0 write-up, links, quotes and releases in the same email are out of scope. The piece is practical: it exists to help a rigorous adopter decide what, if anything, changes for them.

Structure

Part	Role
Launch facts (sizes, pricing, shared specs)	Establishes the object under appraisal and immediately problematizes its headline cost comparison
OpenAI's headline benchmark claim (Agents' Last Exam) vs the SWE-Bench Pro reversal + audit article	Presents the vendor's strongest evidence, then undercuts it with a countervailing benchmark and a suspiciously timed audit
First-party hands-on verdict	Substitutes the author's own early-access experience for benchmark scores
New API features (programmatic tool calling, multi-agent, prompt cache breakpoints, detail:original)	Redirects attention from the model to the platform surface
Pelican-benchmark cost/effort probe + OpenAI demo	Concrete, whimsical illustration of cost-vs-reasoning-effort spread

Key terms

Luna / Terra / Sol
The three sizes of the GPT-5.6 family from smallest to largest, priced (per 1M input/output tokens) at $1/$6, $2.50/$15 and $5/$30 respectively; all share a Feb 16 2026 cutoff, 1M context and 128K max output
Agents' Last Exam
An evaluation of long-running professional workflows across 55 fields; OpenAI's flagship benchmark evidence, on which Sol scores 53.6
SWE-Bench Pro
A self-reported coding benchmark on which Claude Fable 5 (80%) decisively beat GPT-5.6 Sol (64.6%) — and which OpenAI publicly questioned (“~30% of tasks broken”) the day before launch
Token efficiency (“more useful work from every token”) Token Efficiency
OpenAI's framing of GPT-5.6's advance as getting more done per token / lower cost-per-task, not a raw-capability jump
Reasoning tokens
The variable, model-specific hidden tokens spent thinking on a task; because they differ so much between models for the same task, price-per-million-tokens no longer indicates real cost
Programmatic Tool Calling
A new API feature letting the model compose and run JavaScript that orchestrates tool calls; Willison speculates it could bridge MCP tools and full terminal sessions that compose CLI utilities, akin to Anthropic's dynamic filtering
Multi-agent (subagents in the core API) Multi-Agent Orchestration
A new API feature letting the model spin up subagents for parallel, focused work — the sub-agent pattern now baked into the core API rather than the harness
Prompt cache breakpoints
A new API feature bringing Claude-style explicit prompt-cache boundaries to OpenAI (vs OpenAI's default automatic detection, which Willison prefers)

Propositions

explicitOpenAI's GPT-5.6 reached general availability in three sizes — Luna, Terra, Sol (smallest to largest) — all sharing a February 16 2026 knowledge cutoff, a 1M-token context window, and 128,000 max output tokens.conf: high
explicitOpenAI positions long-running agentic performance as GPT-5.6's flagship strength, citing Agents' Last Exam (long-running professional workflows across 55 fields) where Sol sets a new high of 53.6, beating Claude Fable 5 by 13.1 points, with the smaller Terra and Luna also beating Fable at around one-sixteenth the cost.conf: high
explicitHeadline price-per-million-token figures no longer meaningfully indicate real cost, because the number of reasoning tokens a model spends on the same task now varies enormously between models.conf: high
supportsGuest post: AI Inference Is Breaking Unit Economics · p2 — Willison's 'price-per-token no longer tells us much because reasoning tokens differ so much' is a concrete instance of the opposing cost vectors p2 describes (per-token costs falling while reasoning inflates per-query compute).
explicitOpenAI frames GPT-5.6's core advance as token efficiency — getting “more useful work from every token,” i.e. a lower cost-per-task rather than a raw-capability leap.conf: high
supportsDeveloper Newsletter: higher usage limits, dreaming for Claude Managed Agents, and Claude Security · p6 — OpenAI pitching GPT-5.6 as 'more useful work from every token' / lower cost-per-task is a direct example of model improvements now being expressed as efficiency (tokens per task) rather than raw capability.
explicitGPT-5.6's benchmark story is mixed and partly self-serving: Claude Fable 5 decisively beat it on SWE-Bench Pro (80% vs 64.6%), and OpenAI published an article the day before launch estimating ~30% of SWE-Bench Pro tasks are broken.conf: high
supportsWhat OpenAI and Anthropic Think Happens Next With AI · p9 — Supports the premise underlying p9 — that benchmarks are weak, gameable capability signals — by showing a lab curating which benchmark it leads with and discrediting (via a same-week audit article) the one it lost; it is evidence for the premise, not for p9's release-timing claim itself.
explicitIn first-party hands-on use, GPT-5.6 Sol is very competent but has not struck Willison as better than Claude Fable 5 on the complex coding tasks he uses Anthropic's model for.conf: high
supportsGenerative AI in the Real World: Aishwarya Naresh Reganti on Making AI Work in Production · p13 — Hands-on parity between GPT-5.6 Sol and Claude Fable 5 on complex coding is direct evidence that frontier models are converging on capability, per p13.
implicitThe most consequential part of the release is a set of new API primitives — Programmatic Tool Calling, native Multi-agent subagents, explicit prompt-cache breakpoints, and detail:original image handling — several of which adopt patterns Anthropic pioneered.conf: medium
explicitProgrammatic Tool Calling — letting the model compose and run JavaScript that orchestrates tool calls — could bridge the gap between MCP tools and full terminal sessions that compose CLI utilities.conf: medium

Arguments

For a rigorous adopter, GPT-5.6 does not clearly surpass Claude Fable 5, and the release's real significance is cross-lab convergence plus new API primitives rather than a decisive capability leap. strength: medium

Explicit premises:

OpenAI's own headline benchmark (Agents' Last Exam) is countered by a benchmark it lost badly (SWE-Bench Pro: 64.6% vs Fable's 80%).
OpenAI published an audit article discrediting SWE-Bench Pro the day before launch, exactly the benchmark on which it lost.
In hands-on complex coding, GPT-5.6 Sol has not struck Willison as better than Fable so far.
The headline gain is framed as efficiency (work per token / lower cost), not raw capability.
Several new API features explicitly mirror Anthropic mechanisms (dynamic filtering, subagents, prompt caching).

Implicit premises:

Vendor-selected benchmarks are motivated evidence and should be discounted absent independent replication.
First-party hands-on testing on one's own hardest tasks is more decisive for an adoption verdict than aggregate benchmark scores.
A single reviewer's early-access sample on unspecified coding tasks generalizes enough to license a comparative verdict (the shakiest load-bearing assumption).
Cross-lab feature borrowing signals convergence/commoditization of the model surface rather than durable differentiation.
Cost claims stated as vendor 'estimated cost' and round multiples ('roughly one-quarter', 'one-sixteenth') are themselves estimates to be treated cautiously.

Inferences:

From mixed + self-serving benchmarks combined with hands-on parity: benchmark leadership does not equal real-world superiority.
From explicit feature borrowing: the frontier is converging, so the model API surface is where the interesting movement is.

Price-per-million-tokens is no longer a meaningful way to compare model cost. strength: high