OpenAI GPT-5.6 Sol: The Model That Made Itself Cheaper

Picture a factory worker who, in their second week on the job, redesigns the assembly line to run 20 percent cheaper, then hands the savings straight to the customer without being asked. That's roughly what happened inside OpenAI this past August, except the "worker" was GPT-5.6 Sol, and the assembly line was the fleet of GPUs serving billions of tokens a day.

One note before we get into it. Frontier AI models move fast. Labs ship new versions, price changes, and entire model lineups within weeks, sometimes days. Treat the specifics below as accurate as of this writing, and check OpenAI's own pricing and model pages before you commit real budget to any of these numbers.

Here's the two-part story worth telling properly. First, Sol optimized its own inference infrastructure, and that work is what paid for its own price cut. Second, thirteen days after that price cut landed, OpenAI shipped GPT-6 Astra, and Sol went from "the best model OpenAI sells" to "the value pick." Both things are true, and together they explain exactly what kind of model Sol actually is.

GPT-5.6 Sol at a Glance

Let's get the shape of it out of the way first.

Source: OpenAI's GPT-5.6 Sol model page. Nothing in that table is a footnote you can skip past. The pricing row is the whole plot of this article.

What GPT-5.6 Sol Is, and Why It Exists

GPT-5.6 Sol is the flagship of OpenAI's GPT-5.6 family, previewed on June 26, 2026 alongside two smaller siblings, Terra and Luna. Sol was built to be the model you reach for when a task needs OpenAI's strongest available reasoning and agentic coding, without going all the way to a dedicated "everything, at any price" model. Terra and Luna trade some of that ceiling for a lower bill, a decision worth its own post rather than a rehash here (see the tier guide linked further down).

Sol replaced GPT-5.5 as OpenAI's top agentic-coding model, picking up real gains on terminal and tool-use tasks along with a much larger context window. For thirteen days, from its August 21 price cut until GPT-6 Astra's September 3 launch, it was also OpenAI's single most capable model at any price. That window is short, but it's exactly why Sol deserves a post of its own: for two weeks, "OpenAI's best" and "OpenAI's best value" were briefly the same model, and the price cut that made that possible is a genuinely unusual story.

The Model That Fixed Its Own Plumbing

This is the part almost nobody has written up properly, and it deserves the space.

Inside a human-led, human-supervised process, OpenAI's engineers set Sol loose on its own serving stack. Not a one-off experiment, but ongoing production work: Sol autonomously rewrote and optimized production kernels, the low-level code that actually runs inference on a GPU, in Triton and Gluon, OpenAI's own kernel languages. Humans directed the effort and reviewed the output. Sol did the actual rewriting.

That kernel work alone cut the end-to-end cost of serving the model by 20 percent.

Separately, Sol designed and ran hundreds of experiments aimed at improving how efficiently the model generates tokens, while also monitoring live training runs and stepping in when something went wrong. That work pushed token-generation efficiency up by more than 15 percent. OpenAI's own account of the project is here: openai.com/index/gpt-5-6.

Put those two numbers together and you get the reason Sol's price dropped on August 21, 2026: input fell 20 percent to $4 per million tokens, and output fell a third to $20 per million tokens, down from $5 and $30. The efficiency gains funded the cut. Nobody just decided to charge less. The model made itself cheaper to run, and OpenAI passed the savings through.

It's worth being precise about what this is and isn't, because the honest version is the more interesting one anyway. This wasn't a model unilaterally rewriting OpenAI's infrastructure while nobody watched. It was a model doing real, hard optimization work, kernel-level GPU code, inside a process humans set up and supervised the whole way through. That's a genuinely new thing for a shipping product to do. It doesn't need embellishing into a runaway-AI story to be worth writing about; the accurate version is already remarkable.

One more detail worth knowing: the price cut didn't apply everywhere. It covers the pay-as-you-go API, Codex credits, and eligible ChatGPT Work plans. If you're using Sol through a consumer ChatGPT Plus, Pro, or Business subscription, your price didn't move, those plans are flat-rate regardless of which model answers your prompt.

What the Benchmarks Actually Say

We're not going to retype a launch chart here. Every number below is labeled by source, and where two benchmark versions could get confused for each other, we're naming exactly which one we mean.

Start with the table OpenAI, Artificial Analysis, and CodeRabbit each contribute a piece of:

Sol also posts standalone numbers worth knowing, all OpenAI's own reported figures: 92.2% on BrowseComp, 62.6% on OSWorld 2.0, 73.5% on ExploitBench 1, and 64.6% on SWE-Bench Pro.

Now, the trap. Go looking for Sol's Terminal-Bench score yourself and you'll find two very different numbers that are not the same benchmark, despite sharing a name. Terminal-Bench 2.1 is the version behind the 88.8% above (91.9% with Ultra mode switched on, more on that in the pricing section). Terminal-Bench 4.0 is a newer, harder revision of the same benchmark family, and on that version Sol scores 37.3%. Same model, completely different scale, different task set. If you ever see a 4.0 number sitting next to a 2.1 number in the same sentence with no version attached, whoever wrote it either didn't check or didn't care. We're flagging it here because it's the single easiest way to get misled about how good Sol actually is at agentic coding.

Independent testing backs up the general shape of Sol's coding strength, even if exact numbers shift depending on who's running the test. CodeRabbit, the code review company, ran its own long-horizon coding benchmark across 100+ multi-language tasks and found Sol passing 63.7% of them versus Terra's 40.7%. On CodeRabbit's separate code-review benchmark, Sol caught 69 of 99 actionable issues in real pull requests (69.7%) against Terra's 52.5%. Artificial Analysis's own Intelligence Index, a broader reasoning benchmark distinct from its Coding Agent Index above, scores Sol at 47, ranking 14th out of 202 models tracked, well above the 24 median for models in its price bracket.

How Does Sol Compare to Claude Fable 5.1?

Since that's the comparison most readers actually came here for: it depends what you're asking it to do. On Terminal-Bench 2.1, Sol's 88.8% leads Fable-class models, and it does so for under half of Claude Fable 5.1's price ($4/$20 versus $10/$50 per million tokens). But on SWE-Bench Pro, a benchmark that leans harder on complex, multi-file repository patches rather than terminal and tool orchestration, independent comparisons put Claude Fable 5.1 roughly 15 to 16 points ahead of Sol's 64.6%. Different benchmarks measure different skills, and the two models simply aren't optimized for the same one.

What GPT-5.6 Sol Costs

We'll keep this section tight, because a full breakdown of what to pay across OpenAI's entire GPT-5.6 lineup lives in our GPT-5.6 pricing guide, not here.

Sol's own numbers, promotional through November 21, 2026: $4 per million input tokens, $0.40 for cached input, $20 per million output tokens. That's down from $5/$30 before the August 21 cut, funded by the efficiency work described above. Terra and Luna, the two other models in the GPT-5.6 family, both undercut Sol further on price at the cost of some capability. We walk through the exact math on all three, including where each one actually earns its keep, in the Luna post linked above.

Two mechanics actually change your bill, and neither is optional detail:

• The long-context cliff. Push a prompt past 272,000 input tokens and the entire request, not just the tokens over the line, bills at 2x input and 1.5x output. This cliff mechanic is the same across the whole GPT-5.6 family, so it's worth checking your average prompt size before assuming the sticker price is what you'll actually pay.

• Fast mode. Runs up to 2.5x standard speed for 2x the standard price, with no change in the model's actual intelligence. It's purely a latency trade, worth it if you're paying for speed, not capability.

• Ultra mode. Spawns four subagents by default, working a task in parallel and merging the results. It's worth +3.1 points on Terminal-Bench 2.1 (88.8% to 91.9%), for roughly 3x the cost of running Sol solo. For most people, that's a bad trade: you're paying triple for a benchmark bump most workloads will never notice. Save it for the rare task where getting it right on the first pass is worth real money.

Reasoning is configurable too. reasoning.mode: "pro" is available, effort runs anywhere from none to max per request, and reasoning now persists across turns instead of resetting each time you send a new message.

If you're trying to work out which of the three GPT-5.6 tiers actually fits your workload, rather than just what Sol costs on its own, that decision lives in our GPT-5.6 tier-selection guide.

Where It's Strong, Where It Falls Short

Where it's strong:

• Terminal and tool-use heavy agentic coding, where Sol's Terminal-Bench 2.1 score leads its own family and holds up against rival flagships.

• Long-context recall. A 91.5% MRCR score at over a million tokens of context is a real, useful number, not a marketing footnote, especially set against Luna's 41.3% on the same task.

• Cost-to-capability ratio. At $4/$20 per million tokens, Sol undercuts both Claude Fable 5.1 and OpenAI's own newer flagship, GPT-6 Astra, by more than half, while still leading its family on most agentic benchmarks.

• Browsing and computer-use tasks. 92.2% on BrowseComp and 62.6% on OSWorld 2.0 are both strong showings for a model that isn't OpenAI's newest anymore.

Where it falls short:

• Repository-level patch work. Independent SWE-Bench Pro comparisons put Claude Fable 5.1 well ahead of Sol's 64.6% here.

• No Realtime, Assistants, fine-tuning, or embeddings endpoints. Sol is scoped to Chat Completions, Responses, and Batch, so if your stack needs any of those other endpoints, it's not an option.

• Ultra mode's cost. Tripling your bill for a 3-point benchmark gain is a rough trade for most production budgets.

• Text-only output. Image comes in, but only text comes out, no image, audio, or video generation.

• It's already the second-tier flagship. GPT-6 Astra now sits above it, so if you need OpenAI's absolute best regardless of cost, Sol isn't that model anymore.

How to Actually Get GPT-5.6 Sol

Sol is live now, not gated behind a waitlist. You can reach it three ways:

• API, pay-as-you-go. Call gpt-5.6-sol through the Chat Completions or Responses endpoint, with support for streaming, structured outputs, function calling, file search, image input, web search, and prompt caching.

• Codex credits, if you're using OpenAI's coding-agent product rather than calling the API directly.

• ChatGPT, on eligible Work plans at the discounted rate. Plus, Pro, and Business subscribers get Sol too, just without the August price cut changing what they pay, since those plans are flat-rate regardless of which model answers a given prompt.

Batch processing is supported for anything that doesn't need a live response, useful if you're running large offline evaluation or data-processing jobs and want to trade latency for a lower bill.

Sol is not deprecated, and OpenAI hasn't announced a sunset date despite Astra now sitting above it in the lineup. If you've already built against Sol, there's no immediate reason to migrate.

Who Should Use Sol, and Who Should Skip It

Use Sol if you're running high-volume coding agents, especially anything terminal- or tool-call-heavy, and cost per task actually shows up in your budget. It's also the right call if you need to hold a very large document, codebase, or conversation in context and retrieve from it reliably. That 91.5% MRCR score is doing real work, not just padding a spec sheet.

If you're pairing Sol with an editor, Cursor and similar AI-first editors can call it directly through the API, and a code-review layer like CodeRabbit is a sensible pairing given how much of Sol's own benchmark story runs through coding tasks. If you're still comparing coding platforms generally rather than just picking a model, our roundup of the AI coding platforms that actually held up in testing is the better starting point than this post.

Skip Sol if you need the single highest ceiling on complex, multi-file repository patches regardless of cost. That's Claude Fable 5.1's strength, not Sol's. Skip it too if your workload needs Realtime, Assistants, fine-tuning, or embeddings, none of which Sol supports. And skip it if raw frontier capability matters more to you than price, since GPT-6 Astra now sits above Sol in OpenAI's own lineup for exactly that reason.

The Verdict

Here's the honest read. Sol is no longer OpenAI's best model. Astra took that title on September 3, 2026, thirteen days after Sol's price cut. But "best" and "right default" aren't the same question, and for most people running real workloads, Sol is the better everyday pick. At $4 input and $20 output per million tokens, it costs 40 percent of what Astra charges, while still leading its own family on nearly every agentic benchmark that matters.

The self-optimization story is the part worth remembering longest. Sol didn't get cheaper because OpenAI decided to discount it. It got cheaper because the model itself found 20 percent of cost sitting in its own kernels and 15 percent more sitting in its own token generation, and OpenAI let it fix both, under supervision, then passed the savings straight through. That's a genuinely new kind of story for a shipping AI product, and it's the reason Sol is worth understanding on its own terms rather than just as "the model before Astra."

The sensible pattern going forward isn't reaching for whichever model launched most recently. It's routing: Sol as your default for coding-heavy, cost-sensitive work, and something newer only for the specific tasks that actually need it. If you're weighing that decision across OpenAI's full GPT-5.6 lineup, start with our tier-selection guide; if it's the raw dollar math you need, the full pricing breakdown is here.