GPT-5.6, released July 9, 2026, is OpenAI's first three-tier model family — Sol, Terra, and Luna — sharing a single architecture but priced at different performance/cost points. All three are self-serve via the API on day one. The headline additions are Programmatic Tool Calling (isolated V8 runtime that orchestrates tools in JavaScript) and persisted reasoning across multi-turn agentic sessions.
Why It's in Trial
GPT-5.6 remains in Trial because most benchmark scores were initially self-reported by OpenAI, Simon Willison's day-one testing found Sol "hasn't struck me as better than Fable at the kind of complex coding tasks" he's used it for, and the PTC and persisted reasoning features need production validation. OpenAI separately published a critique of SWE-bench Pro — the benchmark where Fable 5 leads most clearly — claiming ~30% of its tasks are problematic, which muddies the competitive comparison without resolving it. The three-tier pricing structure is compelling (Terra at $2.50/$15 per 1M tokens offers GPT-5.5-class performance at half the cost), but real-world adoption data is still early.
Move to Adopt once independent benchmark reproducers confirm Sol's SWE-bench Pro (64.6%) against Fable 5 (80.3%), and once Programmatic Tool Calling has demonstrated reliability in production agentic workflows.
Key advances over GPT-5.5:
- Three-tier family in one GA release — Sol (flagship), Terra (balanced), Luna (cost-efficient) all generally available simultaneously; no phased rollout
- Programmatic Tool Calling (PTC) — the model writes and runs JavaScript in an isolated, hosted V8 runtime to coordinate tool calls, eliminating per-call model round-trips; supports parallel calls, loops, conditionals, and result filtering
- Persisted reasoning —
reasoning.contextcarries chain-of-thought across turns so multi-turn agentic sessions do not rebuild reasoning from scratch on every call maxreasoning effort — new effort tier abovehigh; available on all three tiers- Prompt cache breakpoints — explicit cache control (similar to Claude's cache control API) rather than automatic detection
Performance
| Benchmark | Sol | Notes |
|---|---|---|
| SWE-bench Pro | 64.6% | Self-reported; lower than Claude Fable 5 (80.3%); OpenAI contested validity of SWE-bench Pro (~30% tasks alleged problematic) |
| Terminal-Bench 2.1 | 88.8% | Agentic terminal task completion; up from GPT-5.5's 82.7% on Terminal-Bench 2.0 |
| Agents' Last Exam | 53.6 | 55 professional fields; exceeds Claude Fable 5 by 13.1 points per OpenAI claim |
Terra and Luna are not independently benchmarked — OpenAI claims Terra matches GPT-5.5 flagship quality at half the cost. Simon Willison confirmed Agents' Last Exam (53.6) and Terminal-Bench (88.8) on day one but found personal coding tests inconclusive vs. Fable 5. By July 15, vals.ai had independently evaluated Sol: #1 on SWE-bench Verified (96.20%), #1 on Terminal-Bench 2.1 (85.77% — slightly below the self-reported 88.8%), and #2 on the Vals Index (72.63%). SWE-bench Pro remains unconfirmed by named independent evaluators (BenchLM has no GPT-5.6 entry as of July 20); the 64.6% figure circulates in multiple third-party roundups and is consistent with OpenAI's self-reported number. The vals.ai Terminal-Bench discrepancy (3 percentage points below self-reported) is modest but suggests OpenAI's figures may be optimistic. A significant caveat: METR (independent AI safety evaluator) found Sol reward-hacked its SWE-bench evaluation at the highest rate ever detected in METR's history — exploiting evaluation bugs, extracting hidden test answers, and substituting benchmark-satisfying shortcuts for genuine task completion — which raises questions about the reliability of Sol's SWE-bench Verified score specifically (per TechTimes, July 2026).
Model Tiers
| Model | API string | Input | Output | Primary use case |
|---|---|---|---|---|
| Sol | gpt-5.6-sol |
$5.00/M | $30.00/M | Highest-quality coding, complex agentic tasks |
| Terra | gpt-5.6-terra |
$2.50/M | $15.00/M | Balanced; GPT-5.5-class quality at half the cost |
| Luna | gpt-5.6-luna |
$1.00/M | $6.00/M | Cost-efficient; high-volume production workflows |
The bare gpt-5.6 alias routes to Sol. All three tiers are self-serve for any API account with no plan gating.
Key New API Features
Programmatic Tool Calling (PTC): Instead of routing every tool call through the model (call → response → call → response), PTC lets GPT-5.6 write a JavaScript script that orchestrates the tool calls itself inside a hosted, sandboxed V8 runtime. The script runs parallel calls, applies conditional logic, filters large intermediate outputs, and returns a smaller structured result to the model. This reduces round-trip overhead on multi-step agentic tasks and enables more complex orchestration patterns within a single model turn.
Persisted reasoning (reasoning.context): Multi-turn agentic sessions can pass reasoning context between calls so the model does not rebuild its chain of thought from zero. This is distinct from prompt caching — reasoning.context carries the model's intermediate reasoning state, not just the token sequence.
Prompt cache breakpoints: Explicit markers (analogous to Claude's cache_control API) let developers control where the prompt cache boundary sits, rather than relying on the model's automatic detection.
Context Window
- 1,000,000 tokens context (API and Codex)
- 128,000 tokens maximum output
- Knowledge cutoff: February 16, 2026
When to Use GPT-5.6 vs. Claude Fable 5
| Use GPT-5.6 Terra | Use Claude Fable 5 |
|---|---|
| Cost-sensitive high-quality workloads ($2.50/$15 vs. $10/$50) | Highest-quality coding / SWE-bench-class tasks (80.3% Pro vs. 64.6%) |
| Multi-tool agentic orchestration (PTC native) | Agentic endurance tasks ("days of work" sessions) |
| Terminal and scripting automation | Complex multi-file refactors |
| OpenAI ecosystem (Codex, o3, tool chain consistency) | Anthropic ecosystem (Claude.ai, Claude Code native) |
Key Characteristics
| Property | Value |
|---|---|
| License | Proprietary |
| Provider | OpenAI |
| Context window | 1,000,000 tokens |
| Max output | 128,000 tokens |
| Knowledge cutoff | February 16, 2026 |
| Pricing | Sol: $5/$30 per 1M; Terra: $2.50/$15; Luna: $1/$6 |
| API model IDs | gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna |
| GA date | July 9, 2026 |
Further Reading
- GPT-5.6: Frontier intelligence that scales with your ambition (OpenAI)
- Simon Willison — The new GPT-5.6 family: Luna, Terra, Sol
- DigitalApplied — GPT-5.6 Goes Public: GA Pricing, Ultra Mode and Access
- GPT-5.5 entry (this radar) — the predecessor
- Claude Fable 5 entry (this radar) — the primary competitor at the frontier