GPT-5.6 vs GPT-5.5: Upgrade and API Migration Guide

Short answer: move new, difficult, and cost-sensitive API workloads toward GPT-5.6, but do not replace every GPT-5.5 route with Sol. Use GPT-5.6 Sol for the hardest agentic coding and knowledge work, trial Terra first for typical GPT-5.5 production traffic, and use Luna only where high volume and lower cost matter more than peak capability. Keep GPT-5.5 temporarily when snapshot stability, legacy cache economics, or an unpassed regression suite matters more than the new features.

GPT-5.6 is generally available across ChatGPT, Codex, and the API, while GPT-5.5 remains available and is not listed for deprecation. For a model-tier overview, use GPT-5.6 Sol vs Terra vs Luna. For the billing boundary between app subscriptions and API tokens, use ChatGPT Subscription vs OpenAI API Pricing. Subscription access is a separate product decision: use ChatGPT pricing, the Codex profile, and Codex pricing instead of treating API token rates as a ChatGPT or Codex plan table.

Recommendation by workload

The main routing change is that GPT-5.6 is a family, not one universal successor configuration. The gpt-5.6 alias routes to gpt-5.6-sol. That is convenient for capability-first experimentation, but it is a poor cost-control mechanism if a request should run on Terra or Luna. Production routers should select gpt-5.6-sol, gpt-5.6-terra, or gpt-5.6-luna deliberately and reserve the alias for workloads where following the flagship is acceptable.

Changed, unchanged, and uncertain

GPT-5.6 Sol has a February 2026 knowledge cutoff, compared with December 2025 for GPT-5.5. That can help on knowledge work whose answer depends on more recent training data, but it is not a replacement for web search, retrieval, or source verification. It also does not prove that Sol will outperform a mature GPT-5.5 prompt on a narrow internal task.

The compatibility story is strong but not perfectly drop-in. Both generations support the Responses API, Chat Completions, Batch, streaming, function calling, structured outputs, text and image input, and text output. Both list web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search. Fine-tuning remains unsupported. If an application only changes the model ID and keeps existing tools, the engineering change can be small; production acceptance is still a tuning and evaluation job.

API pricing and cache economics

Those standard rates do not tell the whole story. For inputs above 272,000 tokens, both GPT-5.5 and the GPT-5.6 family use long-context pricing for the full request: input is charged at twice the short-context rate and output at 1.5 times the short-context rate. Batch and Flex are priced at half the standard rate. Priority processing is different: the published short-context Sol rates are $10 input, $1 cached input, $12.50 cache write, and $60 output per million tokens, while GPT-5.5 is $12.50 input, $1.25 cached input, and $75 output with no extra cache-write fee.

Prompt caching is the migration trap. GPT-5.5 uses the earlier automatic-caching contract, charges no additional fee to write a reusable prefix, and supports the legacy 24-hour retention setting. GPT-5.6 can continue using implicit caching, but it also supports explicit breakpoints and a 30-minute minimum cache lifetime. Its reads retain the 90% input discount, while writes cost 1.25 times uncached input.

That means a cache hit is still cheap, but creating caches is no longer free. Log both cache_write_tokens and cached_tokens; group traffic with a stable prompt_cache_key; place explicit breakpoints only after prefixes likely to be reused; and replace legacy prompt_cache_retention handling with prompt_cache_options.ttl where appropriate. Compare the cost of writes with later reads instead of assuming that the unchanged $0.50 Sol cached-read rate makes cache-heavy GPT-5.5 and Sol workloads economically identical.

Context, reasoning, and tool use

GPT-5.6 Sol, Terra, Luna, and GPT-5.5 each publish a 1,050,000-token context window and a 128,000-token maximum output. The GPT-5.6 family therefore offers more routing choice, not more nominal context. Long-context users should test retrieval accuracy, instruction retention, output completeness, and total billed tokens at the lengths they actually use.

Reasoning control expands in GPT-5.6. GPT-5.5 supports none, low, medium, high, and xhigh; GPT-5.6 adds max. Pro is an execution mode selected with reasoning.mode: "pro", not a separate gpt-5.6-pro model slug. OpenAI recommends preserving the current GPT-5.5 effort as the baseline, testing the same setting and one level lower, and reserving the highest settings for tasks where measured quality justifies added latency and tokens.

Programmatic Tool Calling is the most material API tool-use addition. It lets GPT-5.6 run bounded code that coordinates eligible tools and reduces intermediate results before returning them to the model. It is useful for filtering, joining, ranking, deduplication, aggregation, and validation. It is not automatically better when each result changes the next semantic decision, an action needs approval, or native citations and artifacts must be preserved.

Multi-agent support is initially a Responses API beta. Treat it as a second migration after the single-agent route passes. Multi-agent work can reduce wall-clock time on cleanly separable tasks, but it increases orchestration, token accounting, failure handling, and evaluation complexity. Codex ultra is a product setting that coordinates agents; API developers build an ultra-like route through the multi-agent capability rather than by sending ultra as a model ID.

Benchmark claims versus independent evidence

OpenAI's launch tables report broad gains for Sol over GPT-5.5, including 64.6% versus 59.4% on SWE-Bench Pro, 88.8% versus 85.6% on Terminal-Bench 2.1, 52.7% versus 46.9% on Agents' Last Exam, 90.4% versus 84.4% on BrowseComp, and 62.6% versus 47.5% on OSWorld 2.0. These are OpenAI-published results, even when the benchmark itself was created by an outside organization. ToolColumn did not run or reproduce these tests.

The same launch material also shows why the page should not promise a universal win. On OpenAI MRCR's 512K-to-1M slice, Sol is listed at 73.8% and GPT-5.5 at 74.0%. On Toolathlon, Sol is higher than GPT-5.5, while Terra and Luna are lower. Model choice, reasoning effort, harness, token budget, tool configuration, and scoring method can change the ordering.

Independent external evidence is narrower. Irregular's cyber evaluation calls GPT-5.6 Sol only slightly stronger than GPT-5.5 across its suites. Both models solved all 22 medium and hard Atomic tasks at least once, their category success rates were close, and Irregular says Sol's cost per successful FrontierCyber challenge was similar or slightly higher. That supports a cyber edge in some long-horizon work, not a blanket migration claim.

METR attempted an external software-task time-horizon evaluation but explicitly declined to treat its estimates as robust because detected evaluation exploitation changed the result dramatically; it also did not provide a matched GPT-5.5 baseline. The public verified Terminal-Bench 2.1 leaderboard lists GPT-5.5 results but does not yet list GPT-5.6. OpenAI's Terminal-Bench number should therefore remain labeled as vendor-reported until a comparable public entry appears. These sources are evidence about particular harnesses, not substitutes for application-specific evaluation.

Migration plan and temporary reasons to stay

Start with a frozen GPT-5.5 baseline. Record the exact model or snapshot, reasoning effort, prompt, tools, cache settings, latency class, and representative requests. Grade task success, final-answer completeness, required evidence, tool-call correctness, total tokens, latency, cache reads and writes, and end-to-end cost. Do not count fewer calls or shorter output as an improvement if the user-visible result loses required work.

Shadow the intended GPT-5.6 tier rather than sending every request to Sol. Test Terra first for normal GPT-5.5 traffic, Sol for tasks that fail the quality gate or need the new capabilities, and Luna for high-volume routes with a clear minimum-quality threshold. For each tier, test the current reasoning effort and one level lower. Test max or Pro only on a small quality-first set.

Keep existing prompts initially. OpenAI says GPT-5.5 prompting guidance still applies to GPT-5.6. After the baseline run, remove accumulated instructions or examples only when an evaluation shows they are redundant. Avoid a blanket instruction such as "be concise"; GPT-5.6 is already more compressed and can interpret that wording as permission to omit required parts of an artifact.

Treat cache migration as its own release slice. Measure hit rate and write volume, then decide whether implicit mode is enough or explicit breakpoints improve reuse. Treat Programmatic Tool Calling, persisted reasoning, Pro mode, and multi-agent as later feature migrations with their own handlers, telemetry, budgets, and rollback switches.

Stay temporarily on GPT-5.5 when a dated snapshot is required, free cache writes or 24-hour retention materially improve economics, the GPT-5.6 route does not pass your regression set, or stronger dual-use safeguards add unacceptable false positives or latency. Also stay when the workload gets no measurable quality gain: GPT-5.5 is still available, the published capacity ceiling is unchanged, and a newer family name is not itself a business case.

The migration effort is low for a controlled model-ID trial, medium for a production replacement with routing, cache, prompt, and evaluation changes, and high when the project also adopts Programmatic Tool Calling, persisted reasoning, Pro mode, or multi-agent orchestration. The practical default is a staged router: Terra for the common case, Sol for measured hard cases, Luna for qualified volume, and GPT-5.5 as the temporary rollback until the new routes prove themselves.