Article Overview
The GPT-5.6 family that launched in limited preview last month is now available to every ChatGPT user and every API developer. But the general availability announcement is not just a rollout — it comes with two additions that were not in the original preview.
The first is GPT-5.6 Sol Mini: a distilled version of Sol that costs $0.50 per million input tokens and $3 per million output tokens, carries a one-million-token context window, and substantially outperforms GPT-5.4, GPT-5.5, and GPT-4.1 mini on key benchmarks. OpenAI calls it the best small model they have ever released. At one-tenth the cost of GPT-5.5 with better performance, it is not a hard title to earn.
The second is GPT-5.6 Sol Ultra with enhanced deep research: the subagent-powered Sol variant that already led TerminalBench 2.1 at 91.9% now officially includes the ability to spin up multiple specialized investigative agents in parallel, synthesizing their separate findings into a single comprehensive report. It is available to ChatGPT Pro subscribers and API users.
This article covers the full rollout — what each plan gets, what each model costs, how Sol Mini was built and why it matters, how Sol Ultra deep research works, and what the price-performance trajectory looks like across OpenAI's model history.
Introduction
When GPT-5.6 launched in limited preview, the story was about three things: a new naming system separating generation from capability tier, a benchmark lead in coding, and a government-coordinated rollout. The limited availability meant most developers and users were reading about it rather than using it.
That changes on August 7, 2026. GPT-5.6 Sol, Terra, and Luna are now available to every ChatGPT user and through the standard API. The family is also larger than it was in preview. Sol Mini and Sol Ultra are now official, priced, and accessible — and Sol Mini in particular changes the economics of using frontier-generation AI for high-volume applications in ways that the original three models did not.
The most important number in this announcement is $0.50 per million input tokens. That is what Sol Mini costs. It is also, according to OpenAI's benchmarks, more capable than GPT-5.5 — a model that cost significantly more. The cost of getting GPT-5.5-class performance has dropped from roughly $10 per million input tokens in 2025 to fifty cents in 2026. That is a 20-fold reduction in twelve months.
The Complete GPT-5.6 Family
| Model | Capability Level | Input Price | Output Price | Context | Best For |
|---|---|---|---|---|---|
| GPT-5.6 Sol Mini | Exceeds GPT-5.5 | $0.50/M | $3/M | 1,000,000 tokens | High-volume, cost-sensitive, latency-critical |
| GPT-5.6 Luna | Fast and affordable | $1/M | $6/M | — | Speed-first applications |
| GPT-5.6 Terra | Balanced | $2.50/M | $15/M | — | Most everyday workloads |
| GPT-5.6 Sol | Most capable standard | $5/M | $30/M | — | Demanding tasks, complex reasoning |
| GPT-5.6 Sol Ultra | Frontier with subagents | Same as Sol | Same as Sol | — | Parallel deep research, hardest tasks |
Who Gets What: ChatGPT Plan Breakdown
The rollout gives every tier of ChatGPT user a meaningful upgrade from what they had before.
ChatGPT Free users receive GPT-5.6 Luna as their standard model — a significant step up from the GPT-4o mini-class models that previously powered the free tier. Free users also get access to Sol Mini for select tasks.
ChatGPT Plus subscribers at $20 per month receive GPT-5.6 Terra as their default, with access to Sol, Sol Mini, and Luna. This means Plus subscribers are using a model that OpenAI positions as competitive with GPT-5.5 at half the inference cost.
ChatGPT Pro subscribers at $200 per month receive GPT-5.6 Sol as their default, with access to the full family including Sol Ultra and its deep research capability.
Sol Mini: The Model That Changes the Economics
GPT-5.6 Sol Mini is a distilled version of Sol — a model trained specifically to replicate Sol's behavior and capabilities in a smaller, faster, and significantly cheaper form. The distillation process transfers the reasoning patterns, instruction-following quality, and task capabilities of the larger model into an architecture that runs at a fraction of the compute cost.
The result is a model that OpenAI describes as the best small model they have released. The benchmark positioning is direct: Sol Mini substantially outperforms GPT-5.4, GPT-5.5, and GPT-4.1 mini on key evaluations. It carries the same one-million-token context window as the full Sol. And it runs at $0.50 per million input tokens and $3 per million output tokens.
The pricing deserves context to appreciate. GPT-5.5 costs around $10 per million input tokens. Sol Mini, which beats it on benchmarks, costs $0.50. For a developer running a million input tokens through their application daily, that difference is $9,500 per day — roughly $3.5 million per year. Sol Mini does not just represent a better small model. It represents the collapse of the price premium that frontier-generation quality used to command.
The one-million-token context window at this price point is also new territory. The combination of near-GPT-5.5 quality, 1M context, and sub-$1 input pricing exists nowhere else in the current model landscape.
For developers currently running GPT-4.1 mini workloads — the previous go-to for cost-sensitive applications — the upgrade path is clear. Sol Mini is more capable, cheaper, and carries a significantly larger context window.
The Price-Performance Trajectory
Stepping back from Sol Mini specifically, the broader pattern in OpenAI's pricing over three years tells a consistent story about where AI economics are heading.
In 2023, getting GPT-4-class reasoning cost around $30 per million input tokens. By 2024, GPT-4o delivered comparable quality for $5 per million. By 2025, GPT-5.5 pushed capability further at roughly $10 per million. Now in 2026, GPT-5.6 Terra delivers GPT-5.5-class performance at $2.50 per million, and Sol Mini delivers near-GPT-5.5 performance at $0.50 per million.
The cost of frontier-equivalent AI has dropped by a factor of roughly 60 in three years, while capability has simultaneously increased. This trajectory — more capable, faster, cheaper, simultaneously — is the dominant story of AI model economics, and it is accelerating rather than slowing.
The practical implication is that applications which were economically unviable with earlier model pricing are now becoming viable. A startup that could not afford to run every customer interaction through a capable AI model in 2024 can do so now at Sol Mini pricing. The market for AI applications expands every time another cost threshold is crossed.
Sol Ultra: Parallel Investigation at Scale
GPT-5.6 Sol Ultra was already in the benchmark data from the preview launch — it scored 91.9% on TerminalBench 2.1, leading all models tested. The general availability announcement makes it explicitly accessible and adds a specific new capability: enhanced deep research through parallel subagents.
The architecture is the same one that drove the TerminalBench lead: rather than working through a complex task sequentially, Sol Ultra spins up multiple specialized agents that investigate different aspects of the problem simultaneously. Each agent works independently on its own thread, then the results are synthesized into a single output.
For deep research tasks specifically, this means a query that would require a single agent to work through sources, angles, and considerations one after another can instead be handled by several agents in parallel — one investigating the technical dimension, one the business context, one the regulatory environment, one the competitive landscape — with their findings combined at the end.
OpenAI describes the use cases as complex research tasks that benefit from parallel investigation: due diligence, market analysis, academic literature synthesis, competitive intelligence, and technical documentation creation. These are tasks where the bottleneck is not the quality of any single investigation but the time required to cover enough ground thoroughly.
Sol Ultra is available to ChatGPT Pro subscribers and through the API at the same pricing as standard Sol. The subagent capability is included rather than priced separately.
API Availability: All Five Model Strings
All five GPT-5.6 family members are available through the API with the following identifiers:
| API Model String | Description |
|---|---|
| gpt-5.6-sol | Full Sol capability |
| gpt-5.6-sol-ultra | Sol with parallel subagents and deep research |
| gpt-5.6-terra | Balanced performance and cost |
| gpt-5.6-luna | Fastest and most affordable |
| gpt-5.6-sol-mini | Distilled Sol at $0.50/M input |
For developers who pinned to GPT-4.1 mini for cost-sensitive workloads, gpt-5.6-sol-mini is the direct replacement candidate: more capable, same price tier, larger context window.
How the Benchmarks Look With the Full Family
With Sol Mini and Sol Ultra now included in evaluations, the benchmark picture rounds out across three key measures.
On AIME 2025, the mathematics competition benchmark, Sol Ultra leads the family by leveraging its parallel agent architecture even for math problems. Sol Mini sits competitively with GPT-5.5 despite costing a fraction of the price — the benchmark that most directly demonstrates the value of the distillation approach.
On LMSYS Chatbot Arena, which measures human preference rather than structured task performance, GPT-5.6 Sol is the new leader in the overall ranking. GPT-5.6 Terra ranks competitively when performance is adjusted for cost — the balanced model earning its positioning at the middle of the price curve.
On SWE-bench Verified, which measures real software engineering capability on genuine open-source problems, Sol Ultra leads all models tested. The parallel subagent architecture that helps Sol Ultra on TerminalBench appears to generalize to software engineering tasks where multiple aspects of a codebase need to be investigated simultaneously.
What This Release Means in Practice
For most developers building production applications, this release is primarily about Sol Mini. The combination of GPT-5.5-beating performance at $0.50 per million input tokens, with a one-million-token context window, closes the gap between capability and affordability that has been the defining tension in AI application economics.
The question that Sol Mini answers directly is: can I afford to make this AI-native? For applications where the answer was previously "not at GPT-5.5 quality," the math has changed.
For enterprises running research-intensive workflows — due diligence, competitive analysis, technical literature review — Sol Ultra's deep research capability represents a qualitative change in what a single AI session can produce. Parallel specialized agents synthesizing separate investigations is not the same as a single model working through the same material sequentially. The output is broader, the time is shorter, and the depth of individual threads is not sacrificed for breadth.
For the average ChatGPT Plus subscriber, Terra as the new default means the everyday experience of using ChatGPT improved without the subscription price changing.
Final Takeaway
GPT-5.6's general availability brings the model family to every user simultaneously — free, Plus, and Pro — with a distribution that gives each tier a meaningful upgrade from what came before.
The standout addition is Sol Mini. A distilled version of Sol that outperforms GPT-5.5 at one-tenth the cost, with a one-million-token context window, is the most significant cost-performance shift since GPT-4o made GPT-4-class quality affordable in 2024. For high-volume applications, it resets the economic calculation that governed which AI applications were viable.
Sol Ultra's deep research capability, now officially accessible and explained, rounds out the family's upper end with a use case — parallel multi-agent investigation — that has been theoretically possible but rarely productized at this quality and accessibility.
The price-performance frontier OpenAI announced in this update's title has genuinely advanced. The numbers — $0.50 per million input tokens for a model that outperforms last year's leading frontier model — are the most direct evidence that trend continues.
