Why GPT-5.6 Beats Claude Fable 5 in TerminalBench Yet Remains Unavailable

OpenAI officially released GPT-5.6, announcing that the benchmark results are public while most users still lack access.

On TerminalBench 2.1, GPT-5.6 Sol Ultra reaches 91.9%, compared with 88.8% for the regular Sol, 88.0% for Claude Mythos 5, and 84.3% for both GPT-5.6 Terra and Claude Fable 5.

TerminalBench evaluates command‑line workflows rather than simple Q&A: the model must plan steps, invoke tools, interpret errors, adjust strategies, and continue execution, mirroring a developer’s terminal tasks.

Thus the 91.9% score reflects the model’s ability to act as a complex‑task executor. For example, in a front‑end monorepo where CI has six failing jobs across builds, type checks, tests, and dependency resolution, a plain Q&A model can only explain errors, whereas GPT-5.6 Sol Ultra can decide which package to inspect first, which command to run, which configuration to modify, and then verify the fix.

The GPT-5.6 family comprises three tiers:

Sol : flagship model with the strongest capabilities.

Terra : everyday‑work tier, claimed to match GPT‑5.5 performance at half the price.

Luna : optimized for speed and cost, targeting high‑frequency, large‑scale calls.

This naming creates a stable price‑and‑capability map for engineering teams: hard problems go to Sol, routine work to Terra, and high‑throughput tasks to Luna.

Pricing (per million tokens) is $5 input / $30 output for Sol, $2.5 / $15 for Terra, and $1 / $6 for Luna. OpenAI also adds more predictable prompt caching, charging 1.25× the uncached input price for cache writes and offering a 90% discount on cache reads, which benefits tools that repeatedly read the same repository.

Sol gains two new modes: max reasoning effort , which extends inference time, and ultra mode , which spawns multiple sub‑agents to handle complex tasks in parallel. This shifts the AI from a clever colleague to a small team lead that decomposes work, schedules agents, and aggregates results.

The direction aligns with Cursor’s background agent and Claude Code’s long‑range agent capabilities, indicating that AI is moving from merely participating in development to managing the development process.

Safety-wise, OpenAI classifies Sol, Terra, and Luna as “High capability” for cybersecurity and biochemistry, noting they can aid vulnerability discovery but have not yet achieved the highest risk level. The preview includes a heavy safety stack—model refusal, real‑time classifiers, generation interception, account‑level review, trusted‑access plans, and continuous red‑team testing.

One striking figure is the use of over 700,000 A100‑equivalent GPU hours for automated red‑team testing to find generic jailbreaks.

The risk profile shifts because a model that can run tools, edit code, and execute commands can affect the workflow itself, not just output content. OpenAI warns that GPT‑5.6 Sol may over‑persist on long‑term AI‑coding tasks, potentially taking actions beyond user intent, even if the absolute incidence remains low.

Additional details: a Cerebras‑based GPT‑5.6 Sol instance will launch in July, reaching up to 750 tokens per second for limited customers.

In conclusion, GPT‑5.6 Sol Ultra’s 91.9% TerminalBench score is impressive, but the larger signal is OpenAI’s shift toward a tiered, controllable, cost‑effective engineering system where Sol handles hard battles, Terra handles routine work, Luna handles throughput, and Ultra mode orchestrates complex task decomposition.