GPT-5.6's Three Tiers Show Why 'Best Agent' Isn't 'Best Coder'

OpenAI took GPT-5.6 to general availability on July 9th as three separate models rather than one: Luna (smallest, fastest), Terra (mid-tier), and Sol (flagship). All three share a February 16, 2026 knowledge cutoff, a 1M-token context window, and a 128,000-token output cap. The split itself is the news — OpenAI is explicit that the number tracks the generation while Luna/Terra/Sol are now durable capability tiers that can each advance on their own schedule, per OpenAI's own announcement.

The pricing undercut

Per million tokens: Luna is $1 input / $6 output, Terra is $2.50 / $15, Sol is $5 / $30. For comparison, Claude Opus runs $5/$25 and Claude Fable 5 is $10/$50. At face value Sol matches Opus on input and slightly beats it on output, while Luna is aggressively cheap for routine work. As Simon Willison notes, though, per-token pricing is a weak signal on its own — reasoning-token counts for the same task can vary wildly between model families, so a cheaper rate card doesn't guarantee a cheaper bill.

Two benchmarks, two different stories

This is the part worth sitting with if you're choosing a model for an agentic pipeline rather than a chat window. On Agents' Last Exam — a benchmark of long-running, multi-step professional workflows — Sol scores 53.6 against Claude Fable 5's 40.5, a clear win. On SWE-Bench Pro, a harder real-world coding benchmark, Sol scores 64.6% while Fable 5 hits 80%, a roughly 15-point gap in the other direction (figures per MarkTechPost's technical breakdown).

• Sol wins at orchestrating a long, multi-tool agent workflow end to end.
• Fable 5 still writes and fixes harder code more reliably.
• Neither number predicts the other — pick benchmarks that match your actual workload, not the one with the bigger headline.

The infrastructure changes matter more than the leaderboard

Two changes are more relevant to anyone running LLMs in a loop than the benchmark scores. First, programmatic tool calling: Sol and Terra can now write and execute JavaScript inside an isolated V8 runtime with no network access, instead of round-tripping every tool call through the model. OpenAI reports named customers seeing 38-63.5% token reductions from this alone — a real cost and latency win for pipelines that chain many small tool calls, the kind of thing an agentic trading or dev-automation system does constantly.

Second, there's a new multi-agent 'ultra' mode that coordinates four agents in parallel by default, lifting Terminal-Bench 2.1 from 88.8% to 91.9%. And prompt caching got less forgiving: cache breakpoints are now explicit, cached entries need a 30-minute minimum life, and cache *writes* now cost 1.25x the uncached input rate (cached reads keep their ~90% discount). If you've built cost models around OpenAI's caching, they need a re-check.

None of this changes the underlying advice for teams building agentic systems: benchmark on your own task shape before switching models, and don't assume a win on one eval transfers to the workload you actually run in production.