GPT-5.6 Sol

GPT-5.6 Sol is OpenAI's flagship AI model in the GPT-5.6 family, generally available on July 9, 2026, across ChatGPT, Codex, and the API. Sol is the top "capability tier" for the hardest coding, agentic, and research work. Pricing is $5 per million input tokens, $0.50 per million cached input tokens, and $30 per million output tokens, on a 1,050,000-token context window with a February 16, 2026 knowledge cutoff. We tested it hands-on through the API. Our score: 8.8 out of 10.

Quick Verdict: 8.8 out of 10

Score: 8.8 out of 10. GPT-5.6 Sol is the strongest agentic-coding model OpenAI has shipped, and the first to lead an independent coding leaderboard while costing meaningfully less per task than the model above it on raw intelligence. In our own hands-on testing it wrote a correct hard algorithm on the first try, reasoned cleanly through a logic puzzle, held a fixed JSON schema without hallucinating, and retrieved a buried fact from a long document in about two seconds. It is the right pick if your work is agentic coding, long-horizon tool use, or research on the OpenAI stack. It is the wrong pick if you need the single highest raw-intelligence score, native audio or image generation, or fine-tuning — none of which Sol offers today.

• Best independent result: number one on the Artificial Analysis Coding Agent Index, ahead of Claude Fable 5 and at lower cost per task.
• New "durable capability tier" naming — Sol, Terra, and Luna are permanent tiers, not model sizes, sitting inside the GPT-5.6 generation.
• Programmatic Tool Calling lets the model write and run JavaScript in an isolated sandbox to orchestrate tools; a new "ultra" reasoning effort spins up multi-agent orchestration.
• Honest data gap: Sol is absent from SWE-bench Verified because OpenAI did not submit it to the independent tracker.
• No free plan, no fine-tuning, and no native audio or image output — image generation is a callable tool, not a built-in modality.

What Is GPT-5.6 Sol?

GPT-5.6 Sol is the flagship member of OpenAI's GPT-5.6 model family, announced in a gated preview to trusted partners on June 26, 2026 and released to the public on July 9, 2026 across ChatGPT, Codex, and the API. Its API model identifier is gpt-5.6-sol, with the short alias gpt-5.6. OpenAI positions it as the tier "for the hardest problems" — complex coding, long-horizon agentic work, research, cybersecurity, science, computer use, and design.

The launch also introduced a new naming system worth understanding, because it changes how you read OpenAI's lineup. The number — 5.6 — is the generation. The names Sol, Terra, and Luna are what OpenAI calls "durable capability tiers": persistent quality levels rather than parameter counts or context sizes. Sol is the flagship tier; Terra and Luna sit below it for lighter or cheaper workloads. This is a deliberate move away from the size-implying "mini" and "nano" suffixes, and it means a future GPT-5.7 Sol would be the flagship of that generation. To be explicit, because the internet keeps getting this wrong: this model is GPT-5.6 Sol, not "GPT-6." OpenAI's announcement calls it a "step change in design judgment" that "sets a new standard for both intelligence and efficiency," per its official introduction post and the model documentation.

Crucially, GPT-5.5 is not deprecated. GPT-5.6 is an addition to the lineup, not a forced migration — teams pinned to GPT-5.5 or GPT-5.4 can stay put. One naming caveat: "GPT-5.6 Sol Pro" exists only as a feature of the ChatGPT Business and Enterprise plans, not as a separately billed API model. There is no sol-pro price on the API rate card, so treat any listing that sells it as a distinct model with suspicion.

How We Tested GPT-5.6 Sol

Because Sol shipped only two days before this review, what follows is early hands-on, not a multi-week production burn-in — we want to be upfront about that. On July 11, 2026 we ran gpt-5.6-sol directly through the Chat Completions API on four representative probes: a hard algorithm, a logic puzzle, a strict structured-output extraction, and a long-context retrieval. These are small, deliberate tests, not a full evaluation suite, and we report exactly what we saw. Last tested: July 11, 2026.

Coding. We asked Sol to implement the median of two sorted arrays in O(log(min(m,n))) time using partition binary search — explicitly not the easier merge approach — plus three assert-based tests including an empty-array edge case. Sol returned a correct, idiomatic implementation in about seven seconds, all three asserts passing, and it used the harder partition method we asked for rather than the merge shortcut. Only 76 of its output tokens were reasoning tokens, so it did not overthink a problem it clearly knew.

Reasoning. On the classic three-mislabeled-boxes puzzle, run at "high" reasoning effort, Sol correctly chose to draw from the box labeled "Apples and Oranges" and laid out the full deductive chain — every label wrong means that box must be single-fruit — in under five seconds. Clean logic, no filler.

Structured output. We asked for extraction into a fixed JSON schema. Sol produced schema-perfect JSON, correctly parsing "about 1.05 million tokens" into 1050000 and prose dollar amounts into 5 and 30. Notably, it refused to name a vendor the source text deliberately withheld, returning "unspecified" instead of guessing "OpenAI" — a good anti-hallucination signal. One operational caveat we hit firsthand: reasoning tokens are billed from the same completion-token budget as the answer, so our first run with a low cap returned an empty completion (finish reason "length") before any JSON emitted. Raising the token budget fixed it. If you wire Sol into production agentic loops, budget output tokens generously above the expected answer length.

Long context. We pasted sixty release notes — roughly 1,700 prompt tokens — with a single critical fact buried near the middle. Sol retrieved the exact detail (a token-rotation cadence changed to 37 days, effective a specific build) plus the owner team, and correctly identified two other notes touching the same subsystem, in about 2.3 seconds. Fast and accurate on a needle-in-haystack task. For the underlying speed claim, the third-party evaluator Artificial Analysis measures Sol at roughly 74.5 tokens per second; our single-digit-second latencies on small tasks were consistent with that.

Net read: on the substance we tested, Sol behaved like a disciplined, fast flagship — strong first-try coding, clean reasoning, tight instruction-following, and low hallucination. The one thing to internalize before deploying is the shared reasoning-and-output token budget.

Key Features

Sol's headline additions over the GPT-5.5 generation are a new top reasoning tier, a code-driven tool-orchestration mode, and a broad native tool stack. Here is what actually matters, verified against OpenAI's model documentation and announcement.

Reasoning effort from low to ultra

Sol extends the reasoning effort ladder well beyond prior GPT-5 models. It now runs from low up through xhigh, then two new levels: "max," and "ultra." The ultra level is the interesting one — it enables multi-agent orchestration, spinning up four sub-agents by default and configurable up to sixteen, so a single request can fan work out and reconcile it internally. In our reasoning test at "high" effort the model spent about 105 reasoning tokens; ultra is a different animal aimed at problems where you want the model to plan, delegate, and verify on its own.

Programmatic Tool Calling

The most genuinely new capability is Programmatic Tool Calling. Instead of emitting one tool call at a time and waiting for each result, Sol can write and execute JavaScript inside an isolated, ephemeral V8 runtime to orchestrate multiple tool calls in code — looping, branching, and aggregating results before returning. It is compatible with Zero Data Retention, so the sandbox does not persist your data. For complex agent workflows that today require a lot of glue code around the model, this pushes orchestration logic into the model itself.

Improved prompt caching

Prompt caching is upgraded: cache writes are billed at 1.25 times the input rate, cache reads land a 90% discount, and the minimum cache lifetime is 30 minutes. For long-running agentic loops that re-send a large stable system prompt or codebase, the read discount is what keeps the economics sane at Sol's $30-per-million output rate. Our guide to input, output, and cached token pricing explains why cached reads matter more than headline input price for most agent workloads.

Native tool stack

Sol ships with a wide set of built-in tools callable directly: web search, file search, image generation, code interpreter, a hosted shell, computer use, MCP (Model Context Protocol) client support, and skills. It also supports function calling, structured outputs, streaming, and the Batch API. Two things to be precise about: image generation is a tool the model can invoke, not a native output modality, and there is no native audio in or out. Sol accepts text and image input and returns text.

Core specifications

The hard numbers, from the model reference: a 1,050,000-token context window, up to 128,000 output tokens, and a knowledge cutoff of February 16, 2026. Modalities are text plus image in, text out. Fine-tuning is not supported. If your stack depends on a tuned variant, Sol will not serve it today, and that is a real limitation for teams with custom-trained production models.

GPT-5.6 Sol Pricing in 2026

Sol uses flat per-token pricing with no context-length surcharge — a welcome simplification over the buried long-context tiers some earlier models carried. We verified every figure below directly on OpenAI's API pricing page on July 11, 2026.

API pricing (per 1M tokens)

There is no context-tier surcharge — the same rate applies whether you send one thousand tokens or a million. Cache writes cost 1.25 times the standard input rate, cache reads are discounted 90%, and the Batch API halves both input and output. For overnight and non-latency-sensitive workloads, Batch mode brings Sol down to $2.50 per million input and $15 per million output, which is a serious discount at volume. Because the numbers read as $5 and $30, they are easy to reason about — no hidden step function above a token threshold.

Two clarifications on plans. First, there is no free API tier for Sol; access on ChatGPT depends on a paid plan, and the API is pay-as-you-go. Second, "GPT-5.6 Sol Pro" is a ChatGPT Business and Enterprise feature, not a separate API SKU — you will not find a Pro rate on the API card. If you are new to how input, output, and cached rates combine into a real per-task cost, our pricing explainer walks through the math.

Benchmarks: Independent vs OpenAI-Reported

This is the section where model reviews usually go wrong, so we are splitting it cleanly. Independent numbers come from third-party evaluators that OpenAI does not control. Self-reported numbers come from OpenAI's own launch materials and are labeled as such. Treat the two groups very differently.

Independent benchmarks (third-party evaluators)

The independent read is nuanced and, we think, more useful than the marketing. On aggregate intelligence, Claude Fable 5 is still one point ahead on the Artificial Analysis Intelligence Index (60 to 59) and leads human-preference voting on LMArena. But on the thing Sol is built for — agentic coding — it takes the number one spot on the Coding Agent Index, and it does so while costing about a third as much per task as the model above it on raw intelligence. The glaring gap is SWE-bench Verified: for a model marketed on coding, Sol's absence from the most-watched independent coding benchmark is a genuine hole, and we would want it filled before treating Sol as the undisputed coding leader.

OpenAI-reported benchmarks (self-reported)

These figures are from OpenAI's own announcement. We report them because they are the vendor's claims, but they are not independently verified, and in one case OpenAI itself flags the benchmark as unreliable.

Read these as directional, not decisive. The SWE-Bench Pro row is the clearest example of why: OpenAI reports Sol at 64.6% while citing Fable 5 at 80%, then in the same breath questions whether that benchmark measures what it claims to. When a vendor publishes a number and simultaneously undercuts it, that is a signal to wait for third-party replication. If you want to understand why SWE-Bench Pro and SWE-bench Verified are not interchangeable scores, we broke that down in this explainer on the two SWE-bench variants. Note also what is absent: there is no reliable, sourced standalone GPQA Diamond, AIME 2026, or MMLU figure for Sol, so we deliberately do not cite one — despite content farms that invent them.

Pros and Cons After Testing

What we liked

• Number one on an independent coding leaderboard. The Artificial Analysis Coding Agent Index puts Sol first, ahead of Fable 5 and at lower cost per task — the most credible single data point in the launch.
• Strong, disciplined hands-on behavior. In our tests it wrote a correct hard algorithm first try, reasoned cleanly, and refused to hallucinate a withheld fact in structured extraction.
• Programmatic Tool Calling is a real capability, not a rename. Writing and running JavaScript in an isolated sandbox to orchestrate tools moves genuine glue-code logic into the model, and it is Zero Data Retention compatible.
• Cost-efficiency relative to the frontier. Roughly a third of the per-task cost of the model one point above it on the Intelligence Index makes Sol the efficient frontier choice for coding-heavy work.
• Flat, legible pricing. No context-length surcharge and a straightforward $5 and $30 rate card, with Batch cutting both in half for non-urgent jobs.

Where it falls short

• Absent from SWE-bench Verified. OpenAI did not submit Sol to the most-watched independent coding benchmark, leaving a real hole under a coding-first pitch.
• Not the raw-intelligence leader. Fable 5 is one point ahead on the Intelligence Index and leads human-preference voting on LMArena; Sol ranks eighth there.
• No fine-tuning. Teams with tuned production variants cannot bring them to Sol.
• No native audio or image generation. Image generation is a callable tool, not a modality, and there is no audio in or out — multimodal-heavy products need other models alongside it.
• The shared token-budget gotcha. Reasoning tokens draw from the same completion budget as the answer, which can silently truncate outputs if you cap tokens too tightly, as we hit in testing.

Real-World Use Cases

Agentic coding on the OpenAI stack

Sol's number one Coding Agent Index result and its Programmatic Tool Calling make it the strongest fit for multi-file edits, dependency resolution, and tool-driven coding through Codex and the Responses API. If you already run ChatGPT and Codex, this is the immediate upgrade path for coding agents. If the term is new to you, our explainer on what separates an agentic coding model from a chatbot is worth a read first.

Long-horizon autonomous agents

The "ultra" reasoning tier with built-in multi-agent orchestration (four sub-agents by default, up to sixteen) targets tasks you want the model to decompose, delegate, and verify on its own. Pair it with MCP to expose your own tools without writing custom call shims.

Research and analysis

With web search and file search as native tools and a 1.05M-token window, Sol suits literature synthesis, competitive research, and document-grounded analysis. Its strong long-context retrieval in our testing supports this directly.

Computer use and browser automation

Native computer use lets Sol drive a browser or OS surface as part of a plan — scheduled extraction, form-filling at scale, headed-browser QA. OpenAI reports 62.6% on OSWorld 2.0, though that figure is self-reported.

Structured data extraction pipelines

Structured outputs plus Sol's tight schema adherence — and its refusal to invent withheld fields in our test — make it a solid choice for extraction into fixed schemas, provided you budget output tokens generously.

Cost-sensitive frontier coding via Batch

For overnight or non-urgent coding and analysis jobs, Batch mode at $2.50 per million input and $15 per million output makes frontier-tier coding affordable at scale.

GPT-5.6 Sol vs Fable 5, Opus 4.8, Gemini 3.1 Pro, and GPT-5.5

The mid-2026 frontier is a tight race, and the right model depends on whether you optimize for raw intelligence, coding, price, or ecosystem. All API prices below are per 1M tokens and verified in our own database.

When to pick which

Pick GPT-5.6 Sol for agentic coding leadership, Programmatic Tool Calling, and long-horizon agents on the OpenAI stack — especially if you are already on Codex and the Responses API. It is the efficient frontier for coding-heavy work.

Pick Claude Fable 5 if you want the single highest raw-intelligence and human-preference scores and can absorb $10 and $50 per million — it leads the Intelligence Index and LMArena, though at double Sol's price.

Pick Claude Opus 4.8 if you want a cheaper Anthropic flagship for agentic coding and computer use at $5 and $25 per million — lower output cost than Sol, with strong multi-agent orchestration of its own.

Pick Gemini 3.1 Pro Preview if price-to-intelligence is your binding constraint at $2 and $12 per million, or if you need Google's multimodal breadth. Grok 4.3 goes even cheaper at $1.25 and $2.50 with a free tier if budget dominates.

Stay on GPT-5.5 if it already meets your needs — it is not deprecated, prices identically, and switching mid-project rarely pays off for marginal gains.

Frequently Asked Questions

How much does GPT-5.6 Sol cost in 2026?

GPT-5.6 Sol costs $5.00 per million input tokens, $0.50 per million cached input tokens, and $30.00 per million output tokens on the standard API tier. Batch mode is 50% off at $2.50 input and $15.00 output per million tokens. Priority processing is 2x standard at $10.00 input and $60.00 output per million tokens. Pricing is flat with no context-length surcharge, and there is no free API tier.

When was GPT-5.6 Sol released?

GPT-5.6 Sol became generally available on July 9, 2026, across ChatGPT, Codex, and the API, following a gated preview to trusted partners that began on June 26, 2026. It is part of the GPT-5.6 family alongside the Terra and Luna capability tiers.

Is GPT-5.6 Sol the same as GPT-6?

No. There is no model called "GPT-6." The correct name is GPT-5.6 Sol. In OpenAI's naming system the number 5.6 is the generation, and Sol is a "durable capability tier" — the flagship tier for the hardest problems — sitting alongside the Terra and Luna tiers within the same GPT-5.6 generation.

What is GPT-5.6 Sol's context window?

GPT-5.6 Sol has a 1,050,000-token context window and can produce up to 128,000 output tokens. Its knowledge cutoff is February 16, 2026. It accepts text and image input and returns text output, with no native audio input or output.

What are Sol, Terra, and Luna in the GPT-5.6 family?

Sol, Terra, and Luna are OpenAI's "durable capability tiers" — persistent quality levels rather than model sizes or context lengths. Sol is the flagship tier for the hardest coding, agentic, and research work. Terra and Luna sit below Sol for lighter or more cost-sensitive workloads within the same GPT-5.6 generation.

How does GPT-5.6 Sol compare to Claude Fable 5?

On the independent Artificial Analysis Intelligence Index, Claude Fable 5 leads by one point (60 to 59) and ranks first on LMArena human-preference voting at 1509 versus Sol's 1486. However, GPT-5.6 Sol ranks number one on the Artificial Analysis Coding Agent Index, ahead of Fable 5, and costs about a third as much per task. Sol is $5 and $30 per million tokens versus Fable 5 at $10 and $50.

What is Programmatic Tool Calling in GPT-5.6 Sol?

Programmatic Tool Calling lets GPT-5.6 Sol write and execute JavaScript inside an isolated, ephemeral V8 runtime to orchestrate multiple tool calls in code — looping, branching, and aggregating results before responding. It is compatible with Zero Data Retention, so the sandbox does not persist your data. It moves orchestration logic that normally lives in your application into the model itself.

What is the "ultra" reasoning effort level?

Ultra is the highest reasoning effort level on GPT-5.6 Sol, above low, xhigh, and max. It enables multi-agent orchestration, spinning up four sub-agents by default and configurable up to sixteen, so a single request can decompose work, delegate it, and reconcile the results internally. It targets the hardest long-horizon problems rather than routine queries.

Does GPT-5.6 Sol support fine-tuning?

No. Fine-tuning is not supported on GPT-5.6 Sol. Teams that depend on custom-trained variants cannot bring them to Sol today and will need to keep those workloads on a fine-tunable model. OpenAI has not announced a timeline for fine-tuning support on Sol.

Is GPT-5.6 Sol on SWE-bench Verified?

No. GPT-5.6 Sol is absent from SWE-bench Verified because OpenAI did not submit it to that independent tracker. This is a notable data gap for a model marketed on coding. OpenAI does self-report a 64.6% score on the separate SWE-Bench Pro benchmark, but it questions that benchmark's validity itself, so those numbers should be treated as directional rather than verified.

Can GPT-5.6 Sol generate images or audio?

Not natively. GPT-5.6 Sol accepts text and image input and returns text output. Image generation is available as a callable tool the model can invoke, not a built-in output modality, and there is no native audio input or output. Products that need native multimodal generation should pair Sol with dedicated image or audio models.

Is GPT-5.6 Sol Pro a separate API model?

No. "GPT-5.6 Sol Pro" exists only as a feature of the ChatGPT Business and Enterprise plans, not as a separately billed API model. There is no Sol Pro rate on the API pricing card. Any listing that sells Sol Pro as a distinct API model with its own price should be treated with suspicion.

Verdict: 8.8 out of 10

GPT-5.6 Sol earns an 8.8 out of 10. It is the most convincing agentic-coding flagship OpenAI has shipped: number one on an independent coding leaderboard, roughly a third of the per-task cost of the model above it on raw intelligence, and — in our own early hands-on — disciplined, fast, and resistant to hallucination. Programmatic Tool Calling and the ultra multi-agent tier are substantive additions, not marketing. What keeps it from a higher score is honest: Sol is absent from SWE-bench Verified, it trails Claude Fable 5 by a point on aggregate intelligence and sits eighth on LMArena, it offers no fine-tuning and no native audio or image generation, and its reasoning tokens quietly share the output budget.

Score breakdown:

• Features: 9.3 out of 10 — Programmatic Tool Calling, ultra multi-agent reasoning, a 1.05M context window, and a broad native tool stack. The gaps are fine-tuning and native audio or image output.
• Ease of Use: 8.6 out of 10 — a clean drop-in on the Chat Completions and Responses APIs, with reasoning effort accepted as expected. The shared reasoning-and-output token budget is the one thing to learn before production.
• Value: 8.2 out of 10 — the efficient frontier for coding at $5 and $30 per million, with Batch halving both. Held back by no free tier and an absolute price identical to GPT-5.5.
• Support: 9.0 out of 10 — OpenAI's model documentation, pricing pages, and API tooling remain best-in-class and clearly versioned.

Final word: buy GPT-5.6 Sol if your work is agentic coding, long-horizon tool use, or research on the OpenAI stack — it is the strongest and most cost-efficient option for that profile in mid-2026. Choose Claude Fable 5 instead if you need the single highest raw-intelligence score, or Gemini 3.1 Pro Preview and Grok 4.3 if price-to-intelligence dominates. And wait for a SWE-bench Verified number before crowning Sol the outright coding champion — the independent Coding Agent Index says it leads, but the most-watched verified benchmark still has an empty seat where Sol should be.

Sources

• OpenAI — Introducing GPT-5.6 (official announcement)
• OpenAI — gpt-5.6-sol model documentation
• OpenAI — API pricing
• Artificial Analysis — Intelligence Index and Coding Agent Index
• LMArena — human-preference Elo leaderboard