GPT-5.6 Sol is the flagship of OpenAI’s GPT-5.6 model family and, on the independent numbers, the strongest system the company has ever shipped. It launched publicly on 9 July 2026 alongside two stablemates, GPT-5.6 Terra and GPT-5.6 Luna, after a turbulent fortnight in which the models were limited to a small preview group. The official GPT-5.6 announcement billed the family as frontier intelligence that scales with your ambition, and a month of real-world use has given that claim genuine substance.
This article is the durable reference guide in our AI models and tools hub. It maps the Sol, Terra and Luna tier scheme, sets out the full API price list including the long-context surcharge developers keep tripping over, examines the benchmarks with honest comparisons against Anthropic’s Claude, and explains exactly where each tier is available today. Every figure here is drawn from a dated, linkable source — artificial intelligence coverage carries enough vendor hype already.
The launch story itself was dramatic: a planned June release, a government request, a restricted preview and a clearance review. We reported each twist as news at the time, and this guide cross-links those pieces rather than re-litigating the politics. What matters for buyers now is the settled picture: GPT-5.6 Sol is out, its pricing has stabilised, and its August update in ChatGPT added genuinely useful controls.
Below you will find three comparison tables, three charts built strictly from published figures, a machine learning benchmark deep-dive running from the Intelligence Index to SWE-bench Pro, and a closing look at Doug and Astra — the two projects OpenAI has already confirmed beyond GPT-5.6 Sol. Let’s begin with what the model actually is.
Table of contents
- What Is GPT-5.6 Sol? Inside OpenAI’s New Flagship
- From Restricted Preview to Public Launch
- GPT-5.6 Sol vs Terra vs Luna: The Tier Scheme Explained
- GPT-5.6 Sol Capabilities and Specifications
- GPT-5.6 Sol Pricing: The Complete API Breakdown
- GPT-5.6 Sol Benchmarks: An Honest Placement
- Availability in ChatGPT: Plans, Defaults and Modes
- Accuracy and Reliability: The August Update
- Safety and the Preparedness Framework
- GPT-5.6 Sol vs GPT-5.5 and the Retiring o-Series
- Cloud Platforms and Third-Party Access
- Which Tier Should You Choose? A Practical Guide
- What Comes Next: Doug, Astra and GPT-6
- Frequently Asked Questions
- References
What Is GPT-5.6 Sol? Inside OpenAI's New Flagship
GPT-5.6 Sol sits at the top of a three-tier family that OpenAI unveiled in late June 2026. Sol is the flagship: the deepest reasoning, the strongest agentic coding and the highest ceiling on complex, long-horizon work. Terra is pitched as the balanced everyday model, while Luna is the most cost-efficient of the three.
The distinction is not marketing garnish. On the Artificial Analysis Intelligence Index the three tiers score 59, 55 and 51 respectively at maximum reasoning settings, so the gap between flagship and budget tier is real but narrower than the price gap implies. Compression is a defining trait of this generation: benchmark round-ups note that the entire family scores above 92% on GPQA Diamond, with only 2.3 points separating Sol from Luna.
That compression changes how you buy. In earlier generations the cheap model was a different species from the flagship; here the tiers feel like the same brain run at three budgets. The rest of this guide quantifies exactly when the flagship earns its premium.
GPT-5.6 Sol at a Glance
Public launch: 9 July 2026 · Context window: 1M tokens (1.05M reported) · Maximum output: 128K tokens · Knowledge cutoff: February 2026 · API price: $5.00 input / $30.00 output per 1M tokens · Artificial Analysis Intelligence Index: 59 · Coding Agent Index: 80 (state of the art)
Those headline specifications hide a few caveats — an effective context cap in one popular tool, and a surcharge once prompts grow very long — which we unpack in the capabilities and pricing sections below. As a one-line summary, though: this is a million-token flagship priced at $5.00 in and $30.00 out, and it currently tops the agentic coding leaderboard.
How GPT-5.6 Sol Fits OpenAI’s New Naming Scheme
With this release OpenAI finally rationalised its famously confusing product names. In the scheme introduced alongside GPT-5.6 Sol, the number identifies the model generation, while Sol, Terra and Luna are durable capability tiers. OpenAI says the tier names will persist across generations and can advance on their own cadence: Sol is the flagship, Terra the balanced everyday model, Luna the lowest-cost option.
The old pattern of unpredictable suffixes — mini, nano, Instant, Pro bolted onto version numbers — gives way to something a buyer can actually reason about. When GPT-6 arrives, expect the same three tier names attached to a new generation number.
One Family, Three Cadences
Because the tiers are decoupled, OpenAI can improve the flagship without touching the cheaper models — which is precisely what happened on 6 August 2026, when GPT-5.6 Sol received a substantial update in ChatGPT while Luna’s big moment was an availability change rather than a capability jump. For teams standardising on the family, that decoupling means tier choice and generation choice are now separate decisions, each with its own upgrade rhythm.
From Restricted Preview to Public Launch
The GPT-5.6 family had one of the strangest launches in OpenAI’s history. The models were first released on 26 June 2026 as a limited preview for what the company described as a small group of trusted partners — the consequence of US government restrictions under the Trump administration that derailed the planned public June rollout.
The June Delay Explained
We covered the postponement as it happened in our news post on OpenAI delaying GPT-5.6 at the administration’s request. The short version: the models were ready, the announcement was made, but general access was held back. VentureBeat reported at the time that Sol, Terra and Luna had been unveiled yet remained accessible only to limited preview partners pending a US government review.
For two weeks the industry debated whether a review of this kind would become the template for future frontier releases. That debate belongs to the news cycle; what matters for this guide is how it resolved.
The Review That Cleared GPT-5.6 Sol
The bottleneck broke in early July. A US Commerce Department review cleared the models for broad access, and GPT-5.6 launched publicly on 9 July 2026 as a complete three-tier family. Our earlier report on the White House review of GPT-5.6 Sol covers the political mechanics — who asked for what, and why — so we will not repeat them here. The practical takeaway is that the clearance was full: no capability was carved out of the public release.
Launch Day: 9 July 2026
When the public launch came, it came all at once: three tiers, API access and ChatGPT integration together. That completeness mattered. Rival launches have dribbled out over months of waitlists; OpenAI shipped the whole family in a day, and the pricing moves that followed within three weeks suggest the company was keen to convert the delayed momentum into adoption as fast as possible.
It is worth pausing on how unusual the sequence was: GPT-5.6 Sol went from restricted preview to fully cleared public flagship in under a fortnight, with no capability withheld. Whatever precedent the review set politically, commercially it barely dented the launch — by early August the model was the default recommendation in most developer conversations we see.
GPT-5.6 Sol vs Terra vs Luna: The Tier Scheme Explained
The fastest way to understand the family is side by side. GPT-5.6 Sol carries the flagship price and the flagship scores; Terra deliberately undercuts the previous generation; Luna is priced like a utility. Here is the tier map at a glance, using current API prices and Artificial Analysis scores at maximum reasoning settings.
| Tier | Position | API price (per 1M tokens) | Intelligence Index | ChatGPT role |
|---|---|---|---|---|
| GPT-5.6 Sol | Flagship — deepest reasoning | $5.00 in / $30.00 out | 59 | Paid plans; Sol Pro mode on Pro and Enterprise |
| GPT-5.6 Terra | Balanced everyday model | $2.00 in / $12.00 out | 55 | Mid-tier option in the line-up |
| GPT-5.6 Luna | Most cost-efficient | $0.20 in / $1.20 out | 51 | Default for Free and Go users since 6 August 2026 |
Read the table’s columns together rather than separately. The Intelligence Index spread from top to bottom is eight points; the input-price spread is 25x. That asymmetry is the tier scheme’s whole argument: GPT-5.6 Sol exists for the work that needs the last eight points, and the other tiers exist so you stop paying flagship rates for work that does not.
GPT-5.6 Sol: The Flagship Tier
Sol is where OpenAI concentrates its frontier capability. It holds the family’s top benchmark scores, it is the base for the specialist GPT-5.6-Cyber model, and it is the only tier with a higher-effort Pro mode in ChatGPT. If your workload involves multi-step agentic tasks, deep code reasoning or documents that genuinely need a million tokens of context, GPT-5.6 Sol is the tier the rest of this guide spends most of its time on.
GPT-5.6 Terra: The Balanced Middle
OpenAI’s pitch for Terra is blunt: performance competitive with GPT-5.5 while being 2x cheaper. Its Intelligence Index score of 55 lands four points behind the flagship at 40% of the input price. For high-volume production workloads that do not need frontier reasoning on every call, Terra is the family’s rational default — and its 20% price cut on 30 July 2026 sharpened that case further.
GPT-5.6 Luna: The Budget Workhorse
Luna is the astonishing one. At $0.20 per 1M input tokens after its 80% price cut, it scores 51 on the Intelligence Index — eight points behind the flagship at 4% of the flagship’s input price. Since 6 August 2026 it has also been the default model for ChatGPT’s Free and Go users with unlimited text chats, making it the tier most people will actually meet first.
GPT-5.6 Sol Capabilities and Specifications
Specifications only matter when they survive contact with real workloads, so this section pairs each headline number with its caveat. The capability story for GPT-5.6 Sol is genuinely strong — million-token context, huge output allowance, measurable efficiency gains — but two of the fine-print details have already caught developers out.
GPT-5.6 Sol Context Window: One Million Tokens
OpenRouter’s listing puts the context window at 1M tokens, and gate.ai’s specification sheet reports 1.05M tokens of context with a 128K maximum output. In practice that means an entire codebase, a legal data room or several novels’ worth of material can ride in a single request. The same 1-million-token support arrived on Amazon Bedrock in August 2026 for all three tiers, so the long-context capability is not confined to OpenAI’s own API.
Knowledge Cutoff and Output Limits
The knowledge cutoff is February 2026 — recent enough to cover the GPT-5.5 era but not, of course, the model’s own launch drama. The 128K output ceiling deserves more attention than it usually gets: it is what lets agentic sessions emit whole refactors or long structured reports in one pass rather than being chunked.
A February 2026 cutoff also means the model knows nothing of its own family’s launch, the price cuts or its competitors’ summer releases. For current-events work, pair GPT-5.6 Sol with retrieval or browsing rather than trusting the weights — a standing rule for any frontier model, but easy to forget when a model otherwise feels this current.
GPT-5.6 Sol Token Efficiency
Sam Altman claims GPT-5.6 Sol is 54% more token-efficient for AI coding tasks than previous versions, and here the independent data backs the vendor. Artificial Analysis measures roughly 15,000 output tokens per Intelligence Index task for Sol — fewer than Claude Opus 4.8, GLM-5.2 or Gemini 3.5 Flash. Efficiency of that kind compounds: fewer output tokens per task multiplies directly against the $30.00 output rate.
The Codex CLI Context Cap
Now the caveat. Despite the advertised 1.05M window, Codex CLI currently caps the model’s effective context at about 258K tokens — cut from 353K — and the GitHub issue tracking it is a live developer complaint. If your workflow depends on genuinely huge prompts, test your actual tool chain rather than trusting the model card; the raw API delivers the full window, but wrappers may not.
GPT-5.6 Sol Pricing: The Complete API Breakdown
Pricing is where most GPT-5.6 Sol decisions are actually made, and it has more moving parts than the headline rate suggests: a long-context surcharge, a speed premium, batch discounts and aggressive caching. Here is the complete picture from OpenAI’s official price list, current as of mid-August 2026.
| Model / mode | Input (per 1M) | Cached input | Output (per 1M) | Notes |
|---|---|---|---|---|
| GPT-5.6 Sol (standard) | $5.00 | $0.50 | $30.00 | Requests up to 272K input tokens |
| GPT-5.6 Sol (long context) | $10.00 | $1.00 | $45.00 | Requests over 272K input tokens |
| GPT-5.6 Sol Fast Mode | 2x standard | — | 2x standard | Up to 2.5x faster; replaced Priority Processing |
| GPT-5.6 Terra | $2.00 | — | $12.00 | Long context $4.00 / $18.00 |
| GPT-5.6 Luna | $0.20 | $0.02 | $1.20 | Long context $0.40 / $1.80 |
| GPT-5.6-Cyber | $12.50 | — | $75.00 | Daybreak Red programme only |
Standard GPT-5.6 Sol API Rates
The baseline is $5.00 per 1M input tokens and $30.00 per 1M output tokens, with cached input at $0.50 per 1M. Notably, that is exactly the price GPT-5.5 still charges — so at the flagship tier, this generation’s capability jump arrived at zero price increase. The chart below makes the family’s input-price spread vivid; the figures are the official rates just listed.
The 272K Long-Context Surcharge
This is the line item developers keep tripping over. Once a request’s input exceeds 272K tokens, billing switches to the long-context rate: $10.00 per 1M input and $45.00 per 1M output for the flagship, with cached input at $1.00. That is roughly 2x on input and 1.5x on output. A team that assumes the million-token window is uniformly priced can see invoices double on its heaviest requests — budget for the threshold, or engineer prompts to stay under it.
GPT-5.6 Sol Fast Mode, Batch, Flex and Caching
Alongside the 30 July price moves OpenAI introduced Fast Mode for the flagship in the API: up to 2.5x faster than standard processing at twice the price, replacing the old Priority Processing scheme. In the other direction, Batch and Flex processing run at 50% of standard rates, prompt-cache reads carry a 90% discount, and cache writes cost 1.25x the input price. Combining caching with batch runs is the difference between a painful bill and a trivial one for repetitive high-volume pipelines.
The 30 July Price Cuts
On 30 July 2026, CNBC reported, OpenAI cut Luna’s API price by 80% — from $1.00/$6.00 to $0.20/$1.20 — and Terra’s by 20%, from $2.50/$15.00 to $2.00/$12.00, while leaving the flagship unchanged at $5.00/$30.00. The signal is easy to read: the flagship holds its premium while the volume tiers chase ubiquity. Anyone who priced a Luna-based product in early July should re-run the numbers; the economics improved fivefold overnight.
For flagship buyers the cuts still matter indirectly. A cheaper Terra and a five-times-cheaper Luna reshape the routing arithmetic around GPT-5.6 Sol: every call your router can demote to a lower tier now saves more, which raises the bar a task must clear before the flagship is worth invoking. Price stability at the top plus falling prices beneath is quietly the best of both worlds.
GPT-5.6 Sol Benchmarks: An Honest Placement
Vendor launch posts always show a model winning; independent numbers are messier and more useful. The honest summary: GPT-5.6 Sol sets the state of the art in agentic coding, sits one point off the overall intelligence lead, and clearly loses at least one major software-engineering benchmark to Anthropic’s flagship. Details below, every figure sourced.
GPT-5.6 Sol on the Intelligence Index
On the Artificial Analysis Intelligence Index at maximum reasoning, GPT-5.6 Sol scores 59 against 60 for Anthropic’s Claude Fable 5 — one point behind the leader — with Terra at 55 and Luna at 51. One point is inside the margin where prompt style and task mix decide which model feels smarter for you.
Coding Agent Index: A New State of the Art
On Artificial Analysis’s Coding Agent Index the story flips decisively. The flagship’s score of 80 at maximum reasoning is a new state of the art, with Terra at 77 and Luna at 75 close behind. For autonomous coding agents — plan, edit, run, iterate — this family is the strongest measured line-up available, and even the $0.20 budget tier sits within five points of the state of the art on this index.
Where GPT-5.6 Sol Loses: SWE-bench Pro
Third-party comparisons do not hand OpenAI every coding crown. On SWE-bench Pro, codingfleet’s comparison puts Claude Opus 5 at 79.2% against 64.6% for GPT-5.6 Sol — a wide gap on a benchmark built from realistic software-engineering issues. Agentic coding harness design and repository-scale bug fixing are evidently different skills, and buyers doing heavy repair-style engineering work should weigh this number seriously.
GPQA Diamond and ARC-AGI-3
Two more data points frame the family’s ceiling. On GPQA Diamond, graduate-level science questions, the entire GPT-5.6 family lands above 92% with a spread of just 2.3 points from Sol to Luna. On ARC-AGI-3, the abstraction-and-reasoning benchmark, the flagship scored 13.3% on the public set under the official harness — rising to 38.3% when OpenAI enabled retained reasoning and compaction. That tripling from two settings is a reminder that harness configuration can matter as much as the model.
Cost per Task: The Efficiency Story
Artificial Analysis also measures what its Intelligence Index actually costs to run: $1.04 per task for Sol at maximum reasoning, $0.55 for Terra and $0.21 for Luna. The chart below plots those figures. Combined with the roughly 15,000 output tokens per task noted earlier, the flagship’s premium looks defensible — and Luna’s efficiency looks absurd, in the best way.
The rounded verdict: GPT-5.6 Sol is the best agentic coding model measured to date, an effective co-leader on general intelligence, and second-best on at least one heavyweight software-engineering benchmark. Anyone claiming a clean sweep — in either direction — is selling something.
Availability in ChatGPT: Plans, Defaults and Modes
The family’s consumer footprint changed substantially on 6 August 2026, when OpenAI updated the flagship in ChatGPT and rearranged who gets which tier by default. If you last looked at the ChatGPT model picker in July, it has moved under you.
GPT-5.6 Sol and the Effort Slider
The August update gave Plus and Pro users an effort slider — on web, mobile and desktop — that lets you choose how much effort ChatGPT spends on a response. It is the first time reasoning depth has been a user-facing dial rather than a hidden routing decision, and it effectively lets GPT-5.6 Sol impersonate a faster, cheaper model on demand while keeping the flagship’s quality ceiling one notch away.
GPT-5.6 Luna Goes Free
From the same date, Luna became the default ChatGPT model for Free and Go users, replacing GPT-5.5 Instant, with unlimited text chats and a Think button rolling out the following week, subject to abuse guardrails. GPT-5.5 Instant is a model we knew well — our earlier analysis of GPT-5.5 Instant’s shopping-intent behaviour examined how it steered commercial queries — and its replacement by a benchmark-competitive free tier is a quiet landmark for the free tier’s quality.
GPT-5.6 Sol Pro Mode
At the top of the ladder, the higher-effort Sol Pro mode is reserved for Pro and Enterprise plans. As of August 2026 the full ChatGPT line-up runs the GPT-5.6 trio at the top with GPT-5.5, GPT-5.5 Pro and GPT-5.4 below — the fallback tier for users who prefer the previous generation’s behaviour. The table in the next comparison section places every current and retiring model in one view.
One practical note for plan-choosers: the effort slider makes the Plus tier considerably more flexible than it was in July, because a single model now spans quick answers and deep reasoning. The argument for Pro increasingly rests on Sol Pro mode and usage headroom rather than on access to a different model.
Accuracy and Reliability: The August Update
Benchmarks measure capability; the August update targeted trustworthiness, and OpenAI published unusually specific numbers about it. For a flagship increasingly used for health, legal and financial questions, these are arguably the release’s most consequential figures.
Fewer Factual Errors
In OpenAI’s internal evaluation, answers containing at least one factual error were 68% less common with the updated GPT-5.6 Sol than with GPT-5.5 Instant. The company’s deployment safety report adds that the update cut factual error rates by roughly 60% across challenging test sets. Independent replication will take time, but a claimed two-thirds reduction in error-bearing answers is a bigger real-world quality jump than a benchmark point.
HealthBench Professional Gains
The same report records a +15.6-point gain on HealthBench Professional, the evaluation built around clinician-graded health conversations. Pair that with the error-rate reduction and the August update reads as a deliberate reliability release: the kind of unglamorous work that decides whether a model can be trusted in production, not just admired on a leaderboard.
There is a wider point here. GPT-5.6 Sol competes for exactly the workloads where a fabricated citation or an invented figure is most expensive, and OpenAI evidently knows it: shipping a reliability-led update four weeks after launch, with published numbers, is the behaviour of a vendor selling into regulated industries.
Safety and the Preparedness Framework
OpenAI now publishes deployment safety classifications alongside its releases, and the GPT-5.6 family’s August filing contains the most consequential designations the company has made. This section also covers the specialist cybersecurity variant those designations frame.
How OpenAI Classifies GPT-5.6 Sol
The deployment safety report published on 6 August 2026 classifies the August updates to GPT-5.6 Sol and Luna as High capability in both the Biological & Chemical and Cybersecurity domains under the Preparedness Framework, and Below High in AI Self-Improvement. High capability triggers the framework’s stronger safeguards; it is a formal acknowledgement that the flagship’s abilities in these domains are operationally significant, not hypothetical.
GPT-5.6-Cyber: The Specialist Spin-Off
On 10 August 2026 OpenAI launched GPT-5.6-Cyber, its first purpose-trained cybersecurity model, built on the Sol base. The capability delta is stark: on OpenAI’s internal Advanced Cybersecurity Completion Rate evaluation the specialist completes 95.0% of requests versus 1.5% for the stock flagship — the difference between a refusal wall and a working tool. It has already been used to find two previously unknown vulnerabilities in Chrome’s V8 JavaScript engine, patched by Google as CVE-2026-15903.
Inside the Daybreak Red Programme
Access is correspondingly locked down. GPT-5.6-Cyber costs $12.50 per 1M input tokens and $75.00 per 1M output, and is available only through the applicant-vetted Daybreak Red programme: identity verification, legal attestations and — from 1 September 2026 for individual accounts — hardware security keys. The pattern is worth noting as precedent: dangerous capability shipped, but behind vetting rather than behind a refusal.
GPT-5.6 Sol vs GPT-5.5 and the Retiring o-Series
A flagship is only as impressive as what it replaces. Here is how the current model relates to the generation before it, and what is leaving the line-up.
| Model | Status, August 2026 | API price (in / out per 1M) |
|---|---|---|
| GPT-5.6 Sol | ChatGPT flagship; effort slider since 6 August | $5.00 / $30.00 |
| GPT-5.6 Terra | Balanced mid-tier | $2.00 / $12.00 |
| GPT-5.6 Luna | Free and Go default since 6 August | $0.20 / $1.20 |
| GPT-5.5 | Fallback tier; April 2026 from-scratch retraining | $5.00 / $30.00 |
| GPT-5.5 Pro | High-effort legacy option | $30.00 / $180.00 |
| o3 | Retiring late August 2026 | $2.00 / $8.00 |
| o3-pro | Legacy, still listed | $20.00 / $80.00 |
| o1 | Legacy, still listed | $15.00 / $60.00 |
| GPT-4o, GPT-4.1, GPT-4.5, o4-mini | Retired | — |
GPT-5.5: The Fallback Tier
GPT-5.5, released on 23 April 2026 as a from-scratch retraining rather than a fine-tune of GPT-5.4, remains in the API at $5.00/$30.00 with a 1M-token context window of its own; GPT-5.5 Pro sits above it at $30.00/$180.00. The striking fact is the price parity: the new flagship costs exactly what GPT-5.5 costs while beating it — OpenAI’s own line is that even mid-tier Terra is competitive with GPT-5.5 at half the price. There is no economic case for new API projects to start on the older generation.
The o-Series Sunset
The reasoning-specialist o-series is winding down. o3 remains listed at $2.00/$8.00 but is scheduled to retire in late August 2026, while o3-pro ($20.00/$80.00) and o1 ($15.00/$60.00) linger on the price list. GPT-4o, GPT-4.1, GPT-4.5 and o4-mini are already retired. Teams still routing traffic to o3 have weeks, not months, to migrate — and Terra at $2.00/$12.00 is the natural landing spot.
The sunset also simplifies mental models: one generation number, three tiers, one fallback generation. Eighteen months ago OpenAI’s price list read like an archaeology dig; by September it will describe fewer, clearer choices, with the GPT-5.6 family carrying essentially all new work.
Cloud Platforms and Third-Party Access
Not everyone consumes AI models through OpenAI directly. The family’s footprint across clouds and aggregators grew quickly in its first month, which matters for procurement teams whose contracts and compliance sit with a hyperscaler.
GPT-5.6 Sol on Amazon Bedrock
In August 2026, GPT-5.6 Sol, Terra and Luna gained 1-million-token context window support on Amazon Bedrock. For AWS-committed organisations this is the headline: full long-context capability without a separate OpenAI relationship, inside existing security and billing boundaries. It also signals that OpenAI’s distribution strategy for this family is breadth-first.
OpenRouter and Aggregators
OpenRouter lists the flagship with its 1M-token context window and February 2026 knowledge cutoff, making it straightforward to A/B the model against rivals through a single API surface. Aggregator access is also the practical hedge against the tooling caveats noted earlier — if one harness caps effective context, another route may not.
Multi-channel availability has a procurement upside too: headline pricing is identical wherever OpenAI sets it, but quotas, data-residency terms and support tiers differ by channel. Run GPT-5.6 Sol through the route your compliance team already trusts, and keep a second route configured — the Codex CLI episode shows how quickly a single channel’s behaviour can change.
Which Tier Should You Choose? A Practical Guide
Tier choice is a budgeting decision disguised as a capability decision. The honest heuristic: default to the cheapest tier that clears your quality bar, and promote individual workloads upward only when the numbers say so.
Best Use Cases for GPT-5.6 Sol
The flagship earns its premium in four situations: agentic coding, where its Coding Agent Index lead is the state of the art; genuinely long context, where the million-token window (mind the 272K surcharge) changes what is possible; high-stakes accuracy, where the August update’s error-rate reductions matter most; and frontier reasoning, where its 59 on the Intelligence Index is one point off the best measured anywhere. If your workload is none of these, you are probably paying over the odds.
When Terra or Luna Is the Smarter Buy
Terra at $2.00/$12.00 delivers GPT-5.5-class performance at half that generation’s price — the rational default for production pipelines. Luna at $0.20/$1.20 and $0.21 per Intelligence Index task is the volume play: classification, extraction, summarisation and chat at costs that round to zero. A sensible architecture routes by task: Luna for volume, Terra for the everyday middle, the flagship for the calls that justify $1.04 a task.
Modelling Your Token Spend
Run the arithmetic before committing: at standard rates a flagship call with 100K input tokens costs $0.50 before output; the same call crossing the 272K threshold bills input at $10.00 per 1M. Caching (90% off reads) and Batch or Flex processing (50% off) can claw much of that back. If you want help designing that routing — or building the evaluation harness that decides which tier clears your quality bar — our AI strategy and intelligent automation teams do exactly this work.
Two more levers deserve a line in any budget model. The effort slider means ChatGPT-side costs are now partly behavioural — users can spend less effort on routine questions — and Fast Mode means latency-critical API paths can be bought at exactly 2x rather than over-provisioned. Neither lever existed in June; both belong in a 2026 cost model.
What Comes Next: Doug, Astra and GPT-6
A buying guide should say how long its advice will hold. OpenAI has, unusually, given us concrete signals — two named projects that frame the flagship’s shelf life.
The Doug Pre-Training Project
On 9 August 2026 OpenAI unveiled Doug, its largest pre-training project to date — explicitly distinct from GPT-6 — expected to launch no later than November 2026. Whatever Doug ships as, its scale makes it the first real test of whether raw pre-training still yields jumps, and its arrival will bracket the current flagship’s reign at roughly four months of unchallenged leadership.
Astra and the Return of Scaling
GPT-6 itself is expected to be Astra, while Doug restarts OpenAI’s base-model scaling after roughly two years in which capability gains came mainly from reinforcement learning and inference-time compute. Read together: GPT-5.6 Sol is the peak of the refinement era, and the next act returns to brute scale. For buyers, nothing announced obsoletes the current family this year — but budget cycles for 2027 should assume a new generation exists.
Our advice therefore stands as written through at least the autumn: build on GPT-5.6 Sol where the flagship case is clear, keep routing flexible beneath it, and treat November’s Doug launch as the next scheduled moment to re-run this guide’s arithmetic.
Frequently Asked Questions
Is GPT-5.6 Sol free to use?
No. The flagship is available on paid ChatGPT plans, with the higher-effort Sol Pro mode reserved for Pro and Enterprise. What Free and Go users get — with unlimited text chats since 6 August 2026 — is GPT-5.6 Luna, which replaced GPT-5.5 Instant as the default model.
How much does GPT-5.6 Sol cost through the API?
The standard rate is $5.00 per 1M input tokens and $30.00 per 1M output, with cached input at $0.50. Requests over 272K input tokens bill at the long-context rate of $10.00/$45.00, and Fast Mode doubles the price for up to 2.5x faster processing. Batch and Flex processing halve the standard rates.
What context window does GPT-5.6 Sol have?
A 1M-token context window — 1.05M by gate.ai’s specification sheet — with a 128K maximum output and a February 2026 knowledge cutoff. Beware tool-level caps: Codex CLI currently limits effective context to about 258K tokens despite the advertised window.
Is GPT-5.6 Sol better than Claude Opus 5?
It depends on the work. GPT-5.6 Sol holds the Coding Agent Index state of the art at 80 and sits one point behind Claude Fable 5 on the Intelligence Index (59 vs 60), but Claude Opus 5 wins SWE-bench Pro convincingly at 79.2% to 64.6%. Test both on your own tasks before standardising.
When will GPT-6 replace GPT-5.6 Sol?
No date is announced. OpenAI’s Doug pre-training project — distinct from GPT-6 — is expected no later than November 2026, and GPT-6 is expected to be Astra. Until then GPT-5.6 Sol remains the flagship, and the 6 August update suggests OpenAI will keep improving it in place.
What happened to GPT-5.5 Instant?
It was replaced as ChatGPT’s default for Free and Go users by GPT-5.6 Luna on 6 August 2026. OpenAI’s internal evaluation found answers with at least one factual error were 68% less common with the updated flagship than with GPT-5.5 Instant, so the changeover is a straight quality upgrade for free users.
What is GPT-5.6-Cyber?
A purpose-trained cybersecurity variant built on the GPT-5.6 Sol base, launched on 10 August 2026. It completes 95.0% of requests on OpenAI’s internal Advanced Cybersecurity Completion Rate evaluation versus 1.5% for the stock model, costs $12.50/$75.00 per 1M tokens, and is available only through the vetted Daybreak Red programme.
How do Batch, Flex and caching discounts work?
Batch and Flex processing run at 50% of standard rates across the family. Prompt-cache reads carry a 90% discount, cache writes cost 1.25x the input price, and cached input for GPT-5.6 Sol is $0.50 per 1M tokens at standard context. Combined, repetitive high-volume pipelines can cut effective costs dramatically.
References
GPT-5.6: Frontier intelligence that scales with your ambition (OpenAI)
Improving GPT-5.6 Sol in ChatGPT and expanding access to GPT-5.6 Luna (OpenAI)
Advancing the price-performance frontier with GPT-5.6 (OpenAI)
OpenAI API pricing (official docs)
GPT-5.6 — August Updates (OpenAI Deployment Safety Hub)
GPT-5.6 benchmarks across Intelligence, Speed and Cost (Artificial Analysis)
GPT-5.6 Sol — API pricing and benchmarks (OpenRouter)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
