As of July 9, 2026, GPT-5.6 Sol, Terra, and Luna are publicly available via the OpenAI API — no waitlist, no preview partner requirement. The US government freeze that blocked general access to the GPT-5.6 family lifted yesterday, and OpenAI opened the floodgates overnight.
This is not a soft launch. If you’re using GPT-5.5 in production today, you have a migration decision in front of you. If you’re on Anthropic and evaluating alternatives, you have a competitive comparison to run. Here’s the map.
The Three Tiers, Priced
| Tier | Input (per 1M) | Output (per 1M) | Terminal-Bench 2.1 | Best for |
|---|---|---|---|---|
| Sol Ultra | TBD (premium) | TBD | 91.9% | Long-horizon agents, max reasoning |
| Sol | $5.00 | $30.00 | 88.8% | Coding, research, complex agents |
| Terra | $2.50 | $15.00 | ~GPT-5.5 parity | Standard production workloads |
| Luna | $1.00 | $6.00 | Not published | Classification, summarization, high-volume |
Sol Ultra pricing has not been officially published. It runs parallel subagents via reasoning_effort: "max" — the extra score (91.9% vs 88.8% on Terminal-Bench 2.1) comes from parallelized decomposition, not just longer chains of thought. The cost premium lands primarily in output tokens.
The Competitive Context
The numbers that matter most aren’t in the table above — they’re in the comparison:
Sol vs Fable 5: Sol costs $5/$30. Claude Fable 5 costs $10/$50. Sol scores 88.8% on Terminal-Bench 2.1; Fable 5 scores 84.3%. Sol is cheaper and higher-scoring on agentic coding. If you’re paying Fable 5 prices for coding-heavy workloads, that equation shifted today.
Terra vs Sonnet 5: Terra costs $2.50/$15. Claude Sonnet 5 is currently at intro pricing ($2/$10 through August 31), then moves to $3/$15 standard. They’re essentially at parity on price — Terra matches Sonnet 5’s performance at a comparable cost, but Sonnet 5’s intro period means Anthropic users have a slight edge until September.
Luna vs Haiku 4.5: Luna at $1/$6 competes directly with Haiku 4.5 in the cheap-and-fast tier. No published Terminal-Bench scores for Luna yet, but OpenAI’s stated positioning is high-volume pipelines where speed and price beat reasoning depth.
What Each Tier Is Actually For
Luna — The High-Volume Layer
Luna is the first OpenAI production model below the $2 input threshold. Use it when:
- Intent classification with binary or categorical outputs
- Lightweight text extraction or entity recognition
- Real-time interface responses where latency matters more than depth
- Summarization of content where you’re summarizing again at a higher tier anyway
- Pre-screening or triage before escalating to Terra or Sol
The mistake to avoid: Luna’s cost makes it tempting to run everything through it first. A misrouted hard task will produce worse output and force a retry at a higher tier, costing more than if you’d sent it to Terra initially. Build a fast classifier (ideally using Luna itself) to gate the escalation, not a “try Luna first and retry on failure” loop.
Terra — The Production Default
Terra is the migration target for almost everything currently on GPT-5.5. OpenAI says it matches GPT-5.5 performance at roughly half the cost. If that holds at your task distribution — and early benchmark data suggests it does — the migration pays for itself.
“Production default” means: customer support, document analysis, internal tools, moderate-complexity code tasks, structured extraction from business documents, chatbot responses that don’t require deep reasoning.
Terra at $2.50/$15 also makes it viable to run full-context conversations at scale in a way GPT-5.5 at $5/$30 was not. The context cost math changes at half the input price.
Sol — The Hard-Task Layer
Sol belongs on the ceiling of your routing stack. This is where you send:
- Long-horizon agentic tasks (multi-step, multi-tool, 30+ minutes of reasoning)
- Large-codebase refactors where context coherence matters across thousands of lines
- Security review, architecture analysis, or any task where an error is expensive to catch downstream
- Research synthesis where you need cross-document reasoning, not just summarization
At $5/$30, Sol is meaningfully cheaper than Fable 5 for the same work — and scores higher on agentic coding benchmarks. For teams currently routing their hardest Anthropic tasks to Fable 5, this is worth a head-to-head eval.
Sol Ultra — When Parallelism Pays
Sol Ultra adds reasoning_effort: "max" plus automatic subagent parallelization. The 91.9% vs 88.8% Terminal-Bench gap comes from tasks where decomposing work into parallel subagents produces a better answer than a longer single chain of thought — typically: large parallel refactors, multi-file analysis, tasks with independent subtasks that don’t depend on each other’s intermediate outputs.
The cost premium is output-token heavy. Sol Ultra will generate significantly more tokens than Sol on the same task. The calculus: if you have a task where the quality improvement justifies the extra cost (e.g., a security audit where a missed finding costs $50K, not $5), Ultra is cheap relative to the alternative. For batch processing where slightly lower quality is acceptable, standard Sol at $30/M output is the answer.
The Migration Decision
If you’re on GPT-5.5 today: Terra is your default migration target. Half the cost, equivalent capability. Do a spot-check eval on your most important workloads, but the burden of proof is on staying with GPT-5.5 at this point.
If you’re on Anthropic for coding: Run a head-to-head between Sol and Fable 5 on your actual codebase tasks. Sol scores higher on Terminal-Bench and costs less. The counterarguments are tool-use stability, compliance requirements, and whether your workflows depend on Anthropic-specific behaviors (MCP integrations, Claude Code’s built-in tooling, etc.).
If you’re building multi-model architectures: Luna/Terra/Sol maps cleanly onto a three-layer stack: preprocessing, core reasoning, hard escalations. The pricing cliff between Luna ($1/$6) and Sol ($5/$30) is steep enough that good routing pays for itself at volume.
If you’re on Sonnet 5 through August 31: The $2/$10 intro pricing is genuinely attractive for another two months. Terra doesn’t beat that. Revisit in September when Sonnet 5 moves to standard pricing.
What Anthropic Still Owns
The competitive comparison above is real, but it’s not complete without noting where Anthropic’s advantages survive the GPT-5.6 public launch:
MCP ecosystem: Claude’s native integration with the Model Context Protocol, Claude Code’s enterprise tooling, and the Anthropic-first MCP server ecosystem mean a lot of builder workflows are entangled with Anthropic’s infrastructure in ways that don’t migrate cleanly.
Safety tuning and refusal behavior: For customer-facing applications where behavioral consistency on edge cases matters, the behavioral gap between models is real. Fable 5’s safety profile and refusal calibration is different from Sol’s. That difference isn’t a benchmark number.
Compliance infrastructure: SOC 2, data processing agreements, US data residency — if your enterprise customers’ procurement teams require these, the model score is secondary.
The GPT-5.6 family is now clearly competitive for coding-heavy and agentic workloads. It is not a wholesale replacement for everything running on Anthropic today.
The Short Version
- Luna: high-volume preprocessing, $1/$6, default for anything you’d call “cheap and fast”
- Terra: production default, $2.50/$15, your GPT-5.5 migration target
- Sol: hard tasks, $5/$30, cheaper and higher-scoring than Fable 5 on agentic coding
- Sol Ultra: parallel decomposable tasks, premium pricing, 91.9% Terminal-Bench — use when the quality delta justifies the output cost
- Fable 5: still relevant for MCP-integrated workflows, compliance-sensitive deployments, and behavioral consistency requirements
The public launch today removes the remaining reason to defer evaluation. The model is available. The pricing is published. The benchmarks are in. Now is the time to run the eval.
Pricing as of July 9, 2026, sourced from OpenAI’s published API pricing page and OpenRouter. Terminal-Bench 2.1 scores from the official leaderboard. Sol Ultra pricing not yet published; cost premium is estimated based on increased output-token generation via subagent parallelization. Verify all pricing against current documentation before making infrastructure decisions.