# GPT-5.6 Sol, Terra, and Luna: Developer Guide and Migration > Source: > Published: 2026-07-21 02:10:30+00:00 OpenAI’s GPT-5.6 family went live on July 9 with three tiers — Sol, Terra, and Luna — alongside two new API modes that change how agentic applications get built. The launch matters less as a benchmark moment and more as a forcing function: three API-level breaking changes shipped with it, and if you have tool-calling code in production, at least one of them probably applies to you. Here is what you need to fix before they do. ## Pick Your Tier GPT-5.6 runs as three separate models with meaningfully different price-to-capability trade-offs. Sol is the flagship at $5/$30 per million input/output tokens, Terra is the balanced production tier at $2.50/$15, and Luna handles high-throughput lightweight tasks at $1/$6. The full picture includes [Claude Sonnet 5, which undercuts Sol on output cost by 3x](https://www.edenai.co/post/claude-sonnet-5-vs-gpt-5-6-sol-vs-gemini-3-1-benchmarks-pricing-which-to-use) and leads the coding-task average despite losing on agentic benchmarks. | Model | Input / 1M | Output / 1M | Best For | |---|---|---|---| | GPT-5.6 Sol | $5.00 | $30.00 | Agents, hard reasoning, Ultra Mode | | GPT-5.6 Terra | $2.50 | $15.00 | Production everyday tasks | | GPT-5.6 Luna | $1.00 | $6.00 | Routing, classification, moderation | | Claude Sonnet 5 | $2.00 | $10.00 | Coding, cost-sensitive agentic work | For most teams: Luna handles intent routing and classification, Terra covers production inference, and Sol earns its cost only on genuinely hard multi-step agentic tasks. If you are primarily running coding agents, Sonnet 5 remains competitive at a fraction of Sol’s output cost. ## Three Breaking Changes to Fix Now These shipped silently alongside the model. None of them is announced as a deprecation alarm — which is exactly why they will catch teams off guard. ### 1. Parallel Tool Dispatch Is Now On by Default Function calling now dispatches multiple tool calls in parallel. If your handler assumes sequential calls — processing one result before the next arrives — it will silently fail in production. There is no error; results simply get dropped. Test your tool-call handlers against concurrent dispatch before you route live traffic to GPT-5.6. This is the breaking change most likely to cause a silent outage. ### 2. `tools.runtime` Replaces `code_interpreter` The legacy `code_interpreter` tool is dropped in favor of a unified `tools.runtime` namespace. Code execution now runs server-side with per-call billing. If your integration references `code_interpreter` , it needs to be migrated. The new namespace also changes how execution calls are billed — audit your cost model alongside the migration to avoid billing surprises. ### 3. JSON Schema 2025-12 Structured outputs switch to the [JSON Schema 2025-12 dialect](https://developers.openai.com/api/docs/guides/tools-programmatic-tool-calling). Legacy `response_format` with `type: "json_object"` stops being guaranteed-shape by Q1 2027. You have time, but updating schemas now avoids a silent regression later. Pin your SDK, validate against the new dialect, and run shadow traffic before cutting over. OpenAI is giving 60–90 days to complete the work. That window is tighter than it looks. ## Programmatic Tool Calling The more interesting new surface is [programmatic tool calling in the Responses API](https://developers.openai.com/api/docs/guides/tools-programmatic-tool-calling). Instead of the standard loop — model emits tool call, your server runs it, result goes back, repeat — the model writes JavaScript that orchestrates multiple tool calls itself. That code runs in an isolated V8 sandbox with no network access and no npm. The model can loop, branch, and combine results without your server shuttling messages. The efficiency gain is real: fewer round trips, fewer tokens, lower latency. To enable it, add the `programmatic_tool_calling` tool and opt your eligible tools in with `allowed_callers` . Your response loop also needs to handle three new item types: `program` , program-issued function calls, and `program_output` . Miss any of them and calls go unprocessed — silently. ## Ultra Mode: When to Use It Ultra Mode is available on Sol only. It spawns four to eight parallel subagents, each working a subtask concurrently, with a coordinator synthesizing the outputs. It is designed for long-horizon work: auditing an entire codebase, producing a 50-page research report, coordinating a multi-step product launch. For a well-defined single task, it is overkill — use standard Sol reasoning or drop to Terra. If your bottleneck is tool-call latency, programmatic tool calling is the right lever. If your bottleneck is deliberation time on a hard problem, Ultra Mode is. They solve different things. ## Where GPT-5.6 Wins (and Where It Does Not) Sol leads on Terminal-Bench 2.1 (88.8% versus Sonnet 5’s 80.4%) and on agentic task averages (92 versus 81.9). [Claude Sonnet 5 leads on the coding task average](https://benchlm.ai/compare/claude-sonnet-5-vs-gpt-5-6-sol) (76.7 versus 64.6) and is significantly cheaper on output. Gemini 3.1 Pro still holds the top WebDev Arena Elo at 1,487. GPT-5.6 Sol is strong — it is not dominant across every axis. One thing to note on benchmarks: OpenAI published an audit on July 8 showing that approximately 30% of SWE-bench Pro’s 731 tasks are fundamentally flawed. [OpenAI retracted its own recommendation to use the benchmark.](https://openai.com/index/gpt-5-6/) Frontier pass rates on SWE-bench Pro climbed from 23.3% to 80.3% in eight months — worth keeping in mind how much of that reflects the yardstick shifting rather than the models improving. DeepSWE is the emerging replacement. ## What to Do This Week The tier decision is the easy part. Sol for complex agentic pipelines, Terra for standard production inference, Luna for routing and classification. If you are cost-sensitive and coding-heavy, benchmark Sonnet 5 before committing to Sol on output costs. The migration is where the work is. Audit tool-call handlers for parallel dispatch, migrate off `code_interpreter` , and update your JSON schemas. Run shadow traffic against the preview API for two weeks before cutting live traffic over. OpenAI’s 60-90 day window will pass faster than expected.