GPT-5.6: OpenAI's Sol, Terra, and Luna take aim at Fable 5

OpenAI released GPT-5.6 on July 9, 2026 — exactly thirty days after Anthropic's Claude Fable 5 — and for the first time OpenAI gave its models names instead of numbers: Sol (flagship, $5 / $30 per million tokens), Terra (balanced, $2.50 / $15), and Luna (fast, $1 / $6). Sol's new Ultra mode posts 91.9% on TerminalBench 2.1 — past Claude Mythos 5's 88.0% — while Sol undercuts Fable 5's $10 / $50 on both sides of the meter. Anthropic held the top of the leaderboards for one month; OpenAI took several of them back at lower prices. Whatever you think of vendor-published benchmarks, a thirty-day leapfrog cycle at falling prices is the best possible market for the people building on these models.

What are Sol, Terra, and Luna?

Three tiers of the same generation, tuned for different points on the cost-speed-capability curve. All three share a ~1M-token context window, 128K max output, a February 16, 2026 knowledge cutoff, and reasoning-effort settings from none to max. The API exposes them as gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna.

• Sol — the flagship, and OpenAI's self-declared best coding model yet. Ships an Ultra mode that works longer and delegates to submodels, on the same rate card as base Sol — you pay for the extra tokens it burns, not a premium per token.
• Terra — the production workhorse: GPT-5.5-level quality at half GPT-5.5's price. OpenAI's positioning is that Terra handles 70–80% of production workloads without noticeable quality loss.
• Luna — the fast tier, creating a new $1 / $6 price floor for a frontier-family model. On OpenAI's numbers it beats Claude Opus 4.8 on TerminalBench 2.1 (82.5% vs 78.9%) at a fifth of the input price.

The naming change matters more than it looks. "5.6" identifies the generation; Sol, Terra, and Luna are durable tier names that can move independently — the same structural shift Anthropic made years ago with Opus, Sonnet, and Haiku, and extended in June with the Mythos class. Both vendors have now landed on the same shape: a tier ladder where the generation number is almost incidental.

How does GPT-5.6 compare with Claude Fable 5?

All figures below are OpenAI-published (the Claude numbers appear in OpenAI's own launch comparisons), so apply the usual marketing-with-receipts discount. TerminalBench 2.1 scores command-line agent workflows — currently the closest thing to a shared yardstick for agentic coding.

Two cells deserve a second look. Terra tying Fable 5 at 84.3% is the sharpest line in OpenAI's deck — a $2.50 / $15 model matching a $10 / $50 model on their chosen benchmark. And the only Claude that beats base Sol is Mythos 5, which is restricted to Project Glasswing cyber-defense partners; among models you can actually buy, Sol leads. Beyond TerminalBench, OpenAI reports Sol at max reasoning sets a new high of 80 on the Artificial Analysis Coding Agent Index — 2.8 points above Fable 5 while using less than half the output tokens — and 53.6 on Agents' Last Exam, 13.1 points clear of Fable 5. Sam Altman's headline claim to CNBC: 54% better token efficiency on agentic coding tasks.

The counterpoint from early testers is that benchmarks are not the whole story. The consistent read is that Fable 5 still has the raw-intelligence edge on the hardest long-horizon problems, while GPT-5.6 is the more predictable daily driver. Dan Shipper of Every put it best: "GPT-5.6 is like a Porsche, Fable is like a warp drive." Where each one lands against the wider field is tracked in our coding-model comparison, which now carries the GPT-5.6 rows.

How much does GPT-5.6 cost?

Cached reads get a 90% discount inside a 30-minute reuse window (cache writes cost 1.25× the input rate), and batch jobs run at up to 50% off. For contrast: Claude Fable 5 is $10 / $50, Opus 4.8 is $5 / $25, Gemini 3.1 Pro is $2 / $12, Grok 4.5 is $15 / $75, and DeepSeek V4 Pro sits at $1.74 / $3.48. Sol matching Opus 4.8's $5 input price while claiming Mythos-adjacent benchmark scores is the aggressive move here — and if the 54% token-efficiency claim holds up in independent testing, the effective per-task gap versus Fable 5 is wider than the sticker difference. Caching mechanics and key handling for all of these APIs are covered in our LLM API overview.

Why did the government see GPT-5.6 before you did?

The strangest thread connecting the summer's two big launches is Washington. Fable 5 shipped June 9, was hit with US export controls on June 12 — after Amazon researchers demonstrated a method for bypassing its safeguards to find software vulnerabilities — and was suspended for all users until the controls lifted June 30. OpenAI visibly learned from that sequence: it previewed GPT-5.6's capabilities to the US government ahead of launch and, at the government's request, ran a limited trusted-partner preview from June 26 before the July 9 public release. Same regulator, two playbooks — Anthropic got the emergency brake, OpenAI negotiated the on-ramp.

The timing pressure runs the other way too. Fable 5's free inclusion on Claude subscription plans — originally set to end June 23, then pushed back by the export-control suspension — finally ended on July 7, two days before Sol became the cheaper flagship. And under OpenAI's Preparedness Framework, all three GPT-5.6 tiers are rated High capability in cybersecurity and bio/chem risk, the same territory that triggered Fable 5's classifier-and-fallback design. The two labs are converging on capability and on safety posture; what differs is who blinked first on price.

Which model should you pick now?

• Token-metered agentic coding — Sol is the value play: half Fable 5's price, higher OpenAI-published agent scores, and the efficiency claim compounds on long runs if it holds.
• Hardest long-horizon problems — Fable 5 keeps the raw-intelligence crown per early testers. If the task is "migrate a 50-million-line codebase", the warp drive still beats the Porsche.
• Production APIs and RAG — Terra at $2.50 / $15 is now the default-tier benchmark every other mid-tier model gets measured against.
• High-volume and latency-sensitive — Luna at $1 / $6, unless you are already tuned around Claude Haiku 4.5 at the same input price.

One honest caveat before you re-platform: every number above is launch-week material, published by the vendor that benefits from it. Sol Ultra bills on the standard Sol rate card but burns meaningfully more tokens per task by design, tier-tuned harnesses (Claude Code around Claude models, Codex around GPT) carry real switching costs that benchmark deltas understate, and independent evals of Sol, Terra, and Luna are only starting to land. If your agent stack is working, the rational move is to benchmark on your own workload during the honeymoon pricing — not to migrate on launch-week numbers.

References

• OpenAI — GPT-5.6 announcement — tiers, pricing, benchmark claims, and availability
• OpenAI — Previewing GPT-5.6 Sol — the June 26 limited preview and the government-requested rollout sequence
• Axios — GPT-5.6 buzz builds — competitive framing against Anthropic and xAI, and the Sol Ultra details
• Decrypt — How GPT-5.6 Sol stacks up — cross-vendor TerminalBench 2.1 and pricing table, including the Fable 5 and Gemini 3.1 cells
• Anthropic — Redeploying Claude Fable 5 — the June 12–30 export-control suspension and July 1 restoration
• OpenAI — GPT-5.6 Preview system card — Preparedness Framework ratings: High capability in cybersecurity and biological/chemical risk