GPT-6 Astra vs GPT-5.6 Sol: Double the Input Price

GPT-6 Astra costs twice as much per input token as GPT-5.6 Sol and 67% more per output token. $10.00 / $50.00 against $5.00 / $30.00. OpenAI shipped it on 3 September 2026, and the upgrade question is unusually simple to state: the price went up, so the output has to be worth more.

TL;DR

• Every rate went up. Input 2x, output 1.67x, cached input 2x. No row is cheaper.
• The gains OpenAI published are agentic, not general: computer use, terminal work, exploit discovery.
• The tier naming is gone. No Sol, Terra or Luna on GPT-6. It is Astra and Astra Pro.
• Context is roughly unchanged, 1.05M to 1.1M, same 128K output ceiling.
• GPT-5.6 Sol is still available and still half the input price.

The Price Comparison

Both figures are the live Ofox catalog on 5 September 2026, where Astra’s rates match OpenAI’s list. The full Astra rate card has the Fast-mode and per-task numbers behind these rows.

An input rate that doubles is a bigger deal than it looks for agent workloads, because agents are input-heavy: every turn resends the system prompt, the tool definitions and the accumulated transcript. Cached input doubling from $0.50 to $1.00 compounds the same way. If your workload is one long conversation rather than many short ones, the input column is your bill.

What the Extra Money Buys

OpenAI’s launch benchmarks cluster in one place, and it is not general chat:

Computer use, terminal automation, site reliability, exploit discovery. That is a coherent story about what this model is for: it is priced as an agent, not as a chat model, and the workloads it was tuned for are the ones that run long, call tools repeatedly and either succeed or fail at a task rather than producing text a human grades.

Two of those numbers carry qualifiers worth keeping attached: ARC-AGI-3 at 99.9% is with an adapter harness, and Humanity’s Last Exam at 57.2% is with tools.

The adapter qualifier is doing heavy lifting. On ARC Prize’s own standard harness the same model scores 62.7%, and the Sol figure printed beside it on the launch scorecard was measured on a different harness again — OpenAI’s own earlier estimate put Sol near 30% under adapter settings. Our review has the full grid and what it means for reading generational claims.

Where It Lands on a Third-Party Index

61 on the Artificial Analysis Intelligence Index at max effort, ranked #8 of 202.

That is a strong placement, and it is worth knowing what sits above it, because the answer is not GPT-5.6: it is Claude Fable 5.1 at 66, Claude Opus 5 at 63 and Muse Spark 1.3 at 62. AA’s index weights general reasoning, long-context retrieval and knowledge, which is a different mix from OpenAI’s agentic selection above.

Within Astra’s own tiers, the effort setting costs more than it returns on that index:

One point for roughly double. Start on high.

The Naming Changed, and Configs Will Notice

GPT-5.6 shipped as Sol, Terra and Luna. GPT-6 ships as Astra and Astra Pro.

That matters operationally more than aesthetically. Anything that builds a model string by pattern, a router that appends a tier suffix, a config template with gpt-5.6-{tier}, a fallback chain that assumes three tiers exist, will not find a GPT-6 equivalent. The API model ID is simply gpt-6-astra.

This is the same class of break that GPT-5.6’s bare model ID caused on gateways, where gpt-5.6 without a tier returned 404 on routes that expected an explicit tier. The lesson repeats: model naming schemes are not stable across generations, and code that generates model strings rather than reading them from a catalog breaks on every launch.

When to Upgrade

• Your workload is computer use or terminal automation. This is the case OpenAI made, with specific numbers behind it.
• You are doing security or SRE work where ExploitBench and SRE-Bench describe your actual tasks.
• A human waits on the response and Fast mode’s up-to-2.5x speed at double the rate is worth it. GPT-5.6 Sol has no equivalent lever.

When to Stay on GPT-5.6 Sol

• General reasoning at volume. Half the input price, and the third-party index does not show Astra leading its price class.
• Input-heavy agent loops on a budget. Both the input and cached-input rates doubled, which hits exactly this shape of workload hardest.
• Cost per token is the binding constraint. openai/gpt-5.6-sol is in the Ofox catalog at $5.00 / $30.00, half Astra’s input rate. openai/gpt-6-astra joined the same catalog on 5 September 2026, so switching between them is a one-string change once you have measured the difference. The tier guide covers choosing between Sol, Terra and Luna if you are staying on this generation.

The Cheaper Question Nobody Asks

If the reason to consider Astra is “we need more capability”, the comparison worth running first is not against GPT-5.6. At $10 / $50, Astra costs exactly what Claude Fable 5.1 costs, and Fable 5.1 scores five points higher on the same third-party index with cache reads at a quarter of the price. Whether that matters depends on whether your workload looks like AA’s index or like OpenAI’s benchmark table, but it is the comparison that actually tests the price.

Sources

• https://artificialanalysis.ai/models/gpt-6-astra
• https://artificialanalysis.ai/models/gpt-6-astra-high
• https://llm-stats.com/models/gpt-6-astra
• https://ofox.io/models/openai/gpt-6-astra
• https://ofox.io/models/openai/gpt-5.6-sol

GPT-6 Astra benchmarks are OpenAI’s launch figures from 3 September 2026. Rates and context lengths for both models were read from the live Ofox /v1/models endpoint on 5 September 2026, the day GPT-6 Astra was listed. Index scores are from Artificial Analysis, read 4 September 2026.