OpenAI's GPT-5.6 arrives as a three-tier family — Sol, Terra, Luna — and beats Claude Fable 5 on its headline agentic benchmark, but the verified feedback so far tells a more mixed story.
OpenAI has opened up GPT-5.6 to the public, and it arrives not as a single model but as a three-tier family — Sol, Terra, and Luna. The rollout began on July 9, 2026, roughly two weeks after the models first shipped as a locked-down preview to about 20 US-government-vetted organizations, once the Commerce Department's Center for AI Standards and Innovation cleared wider access. The obvious question for anyone already on Anthropic's Claude Fable 5: does it actually stack up? The honest answer is "it depends on what you measure."
One generation, three tiers
The headline change is structural. Under GPT-5.6's new naming system, the number identifies the generation while the names identify durable capability tiers that can advance on their own cadence:
- Sol — the flagship, aimed at the hardest problems like complex coding and security research.
- Terra — tuned for high-volume business work: customer support, internal tools, document analysis.
- Luna — the fast, low-cost tier for everyday summarizing, drafting, and routine automation.
Pricing and specs
All three share a 1M-token context window, up to 128,000 output tokens, and a February 16, 2026 knowledge cutoff. Pricing per 1M tokens:
- Sol —
$5input /$30output - Terra —
$2.50input /$15output - Luna —
$1input /$6output
Did it stack up to Fable 5?
On OpenAI's chosen headline benchmark, yes. GPT-5.6 Sol posts 53.6 on Agents' Last Exam — which OpenAI frames as a new high that eclipses Claude Fable 5's adaptive-reasoning score by 13.1 points — and OpenAI says Terra and Luna outperform Fable 5 at roughly one-sixteenth the cost. Sol also edges Fable on the Artificial Analysis Coding Agent Index (80 vs 77.2) and turns in a strong Terminal-Bench 2.1 result (88.8%).
The picture flips on other tests. On SWE-Bench Pro, a demanding real-world coding benchmark, Claude Fable 5 leads clearly at around 80% versus Sol's 64.6%. So the "who's better" answer splits along the workload: OpenAI takes the long-horizon agentic and cost-efficiency crowns, while Anthropic's model holds the edge on hard, verifiable software engineering.
What the verified feedback says so far
Early independent testing echoes that nuance rather than the marketing. Developer and long-time model reviewer Simon Willison, who had early access to Sol, called it "definitely very competent" but noted it "hasn't struck me as better than Fable at the kind of complex coding tasks" he runs against Anthropic's model — i.e., parity, not a clear generational leap. It's early, and broad third-party evaluations are still landing, so treat single-benchmark bragging rights on either side with caution.
Reading between the lines
Here's some speculation, clearly labelled as such: Anthropic has kept extending access to Claude Fable 5 — the window is now pushed out to July 19 — and the timing is hard to ignore. It's tempting to read those repeated extensions as a competitive reflex around GPT-5.6's launch, keeping Fable in front of developers at exactly the moment OpenAI is courting them. To be clear, Anthropic hasn't cited GPT-5.6 as the reason; that connection is our read of the tea leaves, not a stated fact.
A genuinely new capability for developers
Beyond the tiers, GPT-5.6 introduces Programmatic Tool Calling in the Responses API: instead of round-tripping every tool call through the model, it runs model-written JavaScript in an isolated V8 runtime with no network access. Named customers report token reductions of 38% to 63.5% on tool-heavy workloads — a real efficiency story for agent builders, independent of raw benchmark scores.
How to get it
The API exposes all three tiers immediately, and ChatGPT Plus, Pro, and Enterprise users get access to Sol. The tiered structure is designed to let you match model to task — Sol when the problem is hard, Terra for volume, Luna when cost dominates.
Our take
GPT-5.6 is a serious release, and the three-tier framing is the smartest part of it — it turns "which GPT" into a cost-and-capability decision rather than a guess. But "did it beat Fable 5?" doesn't have a clean yes. OpenAI wins the benchmark it led with and the price-performance argument; Anthropic still looks stronger on rigorous coding. If you build agents, the more interesting news might be Programmatic Tool Calling, not the leaderboard.
Source: OpenAI — GPT-5.6. Independent hands-on: Simon Willison.
