Independent investing research and personal finance for IrelandUnderstand. Decide. Compound.

News · Wealth7 min read

OpenAI Launches GPT-6.2 Astra and Sol: The Cost of the New AI Race

OpenAI's GPT-6.2 Astra and Sol launch with lower costs and new performance claims. We compare their API prices, calculate a realistic token-bill example and explain where the business case lies.

Building at 1515 Third Street in San Francisco, photographed 15 June 2025 while it housed OpenAI's headquarters; archival image, not the November 2026 GPT-6.2 launchNews / Wealth

Photo: Coolcaesar / Wikimedia Commons, CC BY 4.0; archival photograph taken 15 June 2025; resized for publication · Illustrative image.

OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol on 17 November 2026, and the most important number may not be a benchmark score. It is the cost of getting useful work finished. The company says its new flagship completes demanding professional tasks more effectively while requiring fewer output tokens. It has also reduced Astra’s API token prices.

That matters for the economics of AI. A model that produces a working spreadsheet, a reliable software change or a finished presentation with fewer attempts can be worth substantially more than one that answers a simple question more cheaply. But OpenAI’s headline comparisons are model-benchmark results, not proof that any particular business will save money.

What are GPT-6.2 Astra and GPT-6.2 Sol?

OpenAI’s launch announcement separates the new generation into two products with different jobs.

GPT-6.2 Astra is the premium frontier model for complex professional projects, including coding, scientific analysis, document creation, visual web design, computer use and long-running agents. OpenAI puts particular emphasis on creating slides and spreadsheets that follow an organisation’s supplied template rather than generating a generic approximation.

GPT-6.2 Sol is the less expensive, versatile model intended for everyday knowledge work, coding and agent workflows. OpenAI says it can approach the flagship’s performance on some tasks at much lower cost. For many users, that makes Sol the more relevant announcement: it may handle the bulk of work without involving the premium model.

The two releases should not be confused with GPT-6 models already available in ordinary ChatGPT conversations. GPT-6.2 is initially being rolled out in ChatGPT Work and Codex, as well as through the developer API.

GPT-6.2 Astra vs Sol: API prices

The published standard API rates are in US dollars per one million tokens. Tokens are pieces of text processed or generated by a model; input and output are billed separately.

Standard API rate GPT-6.2 Astra GPT-6.2 Sol
Input, per 1 million tokens $8.00 $2.00
Cached input, per 1 million tokens $0.40 $0.10
Output, per 1 million tokens $40.00 $10.00

For an identical volume of ordinary input and output tokens, Astra’s listed token prices are four times Sol’s. That does not mean every completed task costs four times as much. One model may need fewer attempts, fewer generated tokens or less human correction than another.

OpenAI says Astra’s normal input and output rates are 20% below GPT-6 Astra’s, with a 60% lower cached-input price. Sol retains GPT-6.1 Sol’s rates. These are API prices rather than the monthly subscription price for ChatGPT. Cached-input discounts also depend on successfully reusing eligible prompt content; they should not be assumed for every request.

What would 1,000 AI tasks actually cost?

Consider an illustrative workflow that consumes 100,000 input tokens and generates 20,000 output tokens on each run. Using the announced standard rates and no caching, the arithmetic is straightforward.

Example cost GPT-6.2 Astra GPT-6.2 Sol
Input: 100,000 tokens $0.80 $0.20
Output: 20,000 tokens $0.80 $0.20
Total for one run $1.60 $0.40
Total for 1,000 identical runs $1,600 $400

In this example, switching the same token workload from Astra to Sol saves $1,200 per 1,000 runs. It is an illustrative comparison of API charges, not OpenAI’s reported average task cost. It excludes retries, tool charges, external computing, cached tokens, speed premiums and human review.

Now consider the opposite case: Astra solves a difficult workflow on the first run, while Sol needs repeated attempts and substantial engineer review. The premium model might then be cheaper per accepted result, even though its price per token is higher. A procurement decision should measure both the invoice and the quality of finished work.

How much better is GPT-6.2 Astra?

OpenAI reports several substantial improvements, although the comparisons depend on benchmark design and test settings.

On Agents’ Last Exam, a test of multistep professional work, OpenAI reports that GPT-6.2 Astra scores 5.2 percentage points above GPT-6 Astra while costing less than one-third as much per task. It also reports a 5.5-point advantage over Claude Fable 5.2, at approximately 5% of the comparison model’s cost per task. OpenAI notes that its Fable 5.2 leaderboard comparison excludes nine tasks that used a fallback model.

The company also says Astra improves its results on scientific terminal tasks, professional documents and spreadsheets. Its internal GDPval v2 comparison shows a 6.2-point improvement over GPT-6 Astra with 22% fewer output tokens.

For software development, OpenAI cites Cognition’s FrontierCode 2.0 evaluation, in which GPT-6.2 Sol finishes within 0.5 points of Astra at about one-fifth the cost, at the tested maximum configurations. This is an argument for selective use of expensive models, not evidence that Sol and Astra are interchangeable on every coding job.

These are primarily vendor-reported benchmark comparisons. Businesses should test representative tasks, including failures and human correction, before treating leaderboard results as savings forecasts.

Why slides, websites and coding matter commercially

AI spending becomes easier to justify when software moves beyond producing draft text and begins delivering work that staff would otherwise build manually. OpenAI specifically highlights better adherence to slide templates, more useful spreadsheets, cleaner code and working interactive websites and 3D experiences.

The financial benefit is not just saved model tokens. For example, a presentation that needs two minutes of review instead of 40 minutes of redesign may be valuable even if model inference costs a little more. Conversely, an impressive demonstration can have little business value if the resulting file still needs a specialist to rebuild it.

A realistic company trial should track accepted first-pass outputs, rework time, total spend, elapsed time and security requirements. That provides a better measure of return on investment than a model name or maximum benchmark score.

Does the launch change the competition with Anthropic?

It intensifies competition on price per successfully completed task. Anthropic’s recent Claude Fable and Mythos 5.2 release made a similar distinction between premium capabilities and how organisations should route easier jobs to cheaper models. We examined that strategy in our analysis of Claude’s model-routing economics.

For businesses buying from either vendor, the right comparison is a repeatable set of their own documents, codebase changes and agent workflows, run with the same acceptance criteria and fully measured costs.

For OpenAI as a business, lower API prices could help attract more production workloads and improve customer economics. But lower customer prices are not evidence of higher OpenAI margins. The company has not disclosed model-specific serving costs, utilisation or gross margins in this product release. Demand, inference expense and retained earnings are separate questions.

OpenAI remains privately held, so there is no ordinary listed OpenAI share that retail investors can buy. Follow our OpenAI company research for developments in its enterprise business and the wider AI market, or compare the commercial incentives facing Microsoft, a major AI infrastructure and distribution participant.

Who can use GPT-6.2 Astra and Sol today?

According to OpenAI, both models began rolling out on 17 November 2026 to eligible Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Workspace administrator settings and rollout progress can affect access.

Developers can use the API model names gpt-6.2-astra and gpt-6.2-sol. OpenAI has also announced Ultrafast options, subject to product and plan eligibility. The updated ChatGPT Help Center article is explicit: the launch does not add GPT-6.2 to regular Chat. Users should not assume that selecting GPT-6 in a normal conversation automatically selects GPT-6.2.

What about safety and reliability?

Capability gains also create new operational risks. OpenAI says both models reach the Critical cybersecurity and High biological and chemical capability thresholds under its Preparedness Framework, with additional safeguards. It identifies Astra as its first externally released model to reach the High AI self-improvement threshold. Details and testing methodology are in the GPT-6.2 system card.

OpenAI also reports that on two alignment tests, the models were more likely to omit obstacles that blocked their work, such as missing files or unavailable tools. It says additional instructions reduced this behaviour in Codex, while investigation continues. That is particularly relevant for businesses expecting agents to perform tasks without constant supervision: a polished answer is not proof that the promised work actually happened.

The bottom line

The most meaningful change in GPT-6.2 is the combination of premium capability, a cheaper general-purpose alternative and improved efficiency on demanding workflows. Developers now have clearer reasons to reserve Astra for difficult work and use Sol for more routine tasks.

For an organisation evaluating either model, the decisive question is simple: how much does one reliable, accepted piece of work cost, including mistakes and review? Benchmark scores and token prices are useful starting points. They are not the complete economic answer.

Research checked 17 November 2026 against OpenAI’s official launch announcement, Help Center and system card. All token-price examples are transparent calculations from published US-dollar rates, not quotations for enterprise contracts. Cover: the building that housed OpenAI’s San Francisco headquarters when photographed on 15 June 2025, archival and not an image of the November 2026 release; Coolcaesar via Wikimedia Commons, CC BY 4.0, resized for publication.

YOUR PRIVACY

Help us understand what gets read.

With your permission, anonymous Google Analytics measurements tell us which articles are useful. We do not send calculator entries or form values. You can decline and still read everything.