OpenAI ships GPT-6.1 Sol with 50% API price cut and cache reads at $0.10

GPT-6.1 Sol lands with a 50% Sol and Luna price cut and a $0.10 cache tier

OpenAI released GPT-6.1 Sol on September 29, 2026, framing the model around agentic coding and computer use. Pricing was cut by 50% for the Sol and Luna lines, and cached input is offered at a 95% discount to standard input, with cache reads priced at $0.10 in the published tier. The headline pricing on input and output is reported at $2 and $10, with near-Astra benchmark scores described as the model’s positioning. For developers, the practical change is twofold: agent-style workloads that re-use the same prompt prefix become cheaper to run, and per-token output pricing is halved against the prior Sol and Luna rates.

DevDay 2026: computer use on the Agents API and cloud-based Codex

At DevDay 2026, the headline workflow additions were computer use for the Agents API and a cloud-based Codex environment. The model tier is built around agentic coding and computer-use tasks, and the Agents API now exposes the same capability surface that previously required separate tooling. A cloud-based Codex shifts the editor from a local process to a hosted one, which changes where state, secrets and build caches live. Developers integrating agents that drive a browser or desktop need to retest the failure modes that computer use introduces, and anyone moving code editing into the cloud will need to revisit where their code is stored and which identities can reach it.

What the cache tier changes for prompt-heavy agent loops

The 95% cache-input discount is the most concrete lever a developer can pull this week. Agent loops that re-issue the same system prompt, tool schema and prior turns on every step benefit directly, because the repeated prefix is read at the cache rate rather than the standard rate. At the published $0.10 cache-read number against a $2 standard input figure, the economics shift toward long-running agents with stable prefixes rather than short one-shot calls. The release note also describes cached input as a 95% discount to standard input, which corroborates the same pricing structure across the Sol launch coverage.

What to test first when migrating from Sol and Luna

1 Sol at near-Astra benchmark performance, which is the bar to reproduce in private evals before switching production traffic. Three test surfaces matter: agentic coding tasks with multi-step tool use, computer-use tasks that drive a real UI, and cached-prompt workloads where the 95% discount should appear as a measurable drop in per-run cost. The Agents API gains the computer-use capability, so any existing agent that previously delegated to a separate tool needs to be re-evaluated against the integrated path.

A cloud-based Codex also changes how a developer interacts with a repo, so anyone planning a switch should pilot a non-critical repository first and check how the hosted environment handles credentials, network egress and build state.

A coding workflow shift that lands alongside the model

OpenAI’s coding surface is moving into the cloud at the same time the model prices are moving down. A cloud-based Codex, paired with the Agents API’s computer use, lets one hosted environment combine model calls, tool execution and editor actions in a single loop. For developers, the trade is control versus integration: local Codex keeps code on the workstation and keeps the editor responsive, while the cloud option consolidates state and agent control in one place. Coverage of the DevDay recap and the Sol release guide both list the cloud-based Codex as a ship item rather than a roadmap item, so teams evaluating it should plan a migration window rather than wait for a future note.

Practical follow-ups before the next pricing note

Three concrete checks are worth running this week. First, replay a representative agent trace against the new $2 input and $0.10 cache-read numbers to estimate the new per-task cost. Second, port one agent to the Agents API’s computer-use path and confirm the failure modes match your expectations around retries and partial UI state. Third, if a cloud-based Codex is in scope, dry-run a non-critical repository through the hosted environment and review identity, network and cache settings before any production code touches it. A cached-input discount of 95% is large enough that small prompt-prefix changes can swing the bill, so track cache hit rate as a first-class metric from day one.