The coding-agent price war, September 2026: what Claude Code, Codex, and Gemini CLI actually cost

On Tuesday, September 22, 2026, Anthropic launched Claude Opus 5.5 and OpenAI launched GPT-6 Sol and GPT-6 Luna about 90 minutes apart. All three models arrived cheaper than their predecessors. For engineering managers budgeting coding-agent spend, the headline cuts are only the start of the arithmetic: how teams pay (per seat vs per token), which model the CLI defaults to, and cache economics decide the bill far more than the sticker price.

All prices below are vendor-published rates, labeled as such. Nothing here was live-tested; this is documentation-based research.

What changed in September

Anthropic shipped Opus 5.5 at $4 per million input tokens and $20 per million output, 20 percent below Opus 5. The bigger move was cache reads: $0.20 per million, down 60 percent from $0.50. Anthropic says cache reads make up the majority of agentic and coding-work costs, and that at default settings Opus 5.5 costs about 40 percent less than Opus 5 on typical workloads. Both claims are Anthropic's, not measured here.

OpenAI answered the same afternoon with GPT-6 Sol at $2 in / $10 out per million tokens (half the price of GPT-5.6 Sol) and GPT-6 Luna at $0.10 in / $0.50 out. Cached input reads are 90 percent off. OpenAI also retired the middle tier: GPT-5.6 Terra got no GPT-6 successor, leaving a three-tier ladder of Astra, Sol, Luna.

The CLIs moved in the same week. Claude Code 2.1.280 (September 22) made Opus 5.5 the default Opus model and, more consequentially for budgets, switched the default model on Pro and Team Standard plans from Sonnet to Opus. Pro and Team Standard users now run the most expensive model by default. Claude Code 2.1.278 (September 19) had already made auto mode default to a server-side classifier, eliminating the classifier-overhead charges that used to add a separate usage line on API and Enterprise sessions. Codex CLI 0.153.x to 0.154.0 (early September) put GPT-6 Astra in the model picker and in Amazon Bedrock catalogs, and 0.153.4 made Astra the bundled default. Gemini CLI hit v0.60.0 on September 15, with v0.61.0 stable on September 23.

The numbers, per million tokens

Every cell in this table is a vendor-published rate as reported at the sources listed above. Google's rates carry an expiry worth putting in a budget model: they double on January 1, 2027. Anthropic and OpenAI published no end date on theirs.

The subscription layer, from Anthropic's own pricing page: Claude Pro is $17 a month billed annually ($200 up front) or $20 monthly and includes Claude Code. Max starts at $100 a month for 5x or 20x Pro's usage per five-hour session. Enterprise is $20 a seat a month billed annually, plus usage at API rates. Claude Code is included in all paid plans and draws from one shared usage pool; when you hit the limit you wait, upgrade, or turn on usage credits billed at standard API rates.

Codex is bundled with ChatGPT plans (Plus, Pro, Business, Enterprise, and Edu have access to Codex with the Sol/Luna models; Astra is in the picker and on Bedrock). Gemini CLI is the outlier: free with a personal Google account at 60 model requests per minute and 1,000 per day, which Google calls the largest free allowance in the category. It is also the only one of the three that is open source (Apache 2.0) and the only one with native multimodal input.

What the evidence supports

The unit-price story checks out across multiple independent outlets: Opus 5.5 is 20 percent cheaper than Opus 5 with a 60 percent cache-read cut; Sol and Luna are half the price of their 5.6 predecessors; Astra holds the $10/$50 flagship tier. The version numbers and default changes are verified against official changelogs and release notes.

The per-task story does not check out as cleanly, and both vendors know it. Anthropic's "40 percent less on typical workloads" comes with two documented qualifiers: the default effort setting changed from high to medium in the same release, and "typical workloads" means Anthropic's workloads. OpenAI's cost-per-task charts for Sol mostly compare against Opus 5, not the Opus 5.5 that shipped the same morning. The one independent datapoint worth noting: OpenAI's own chart puts GPT-6 Sol at $0.27 per task on AutomationBench at xhigh effort versus 11x that cost for Opus 5 at max effort, but that comparison predates Opus 5.5's price cut.

Cache economics remain the real lever, and they cut in Anthropic's favor for now. Opus 5.5's cached read costs 5 percent of a fresh input token; Astra's costs 10 percent. In agentic coding work, where each turn re-sends the full context, that ratio decides more of the bill than the headline input price.

The decisions this informs

If your team runs on Claude subscriptions, the default change matters more than the price cut. Pro and Team Standard now default to Opus, which is better output but a heavier draw on the same five-hour limits. Anthropic raised those limits with the launch, but the practical move is to pin Sonnet for routine work in your managed settings or model defaults and reserve Opus 5.5 for the hard tasks.

If you pay per token, route by task. Luna at $0.10/$0.50 is the cheapest volume tier from any frontier lab; Sol at $2/$10 undercuts Opus 5.5's $4/$20 on list price for mid-tier reasoning. Opus 5.5's cache-read price ($0.20) is the best in the table for long sessions that re-read the same context.

If budget certainty is the priority, Gemini's free tier is the only path that costs nothing, and its paid rates are the lowest per token. The catch is printed on Google's own pricing page: those rates double on January 1, 2027. Model any commitment past December at the higher numbers.

What remains unknown

Actual per-task cost across real workloads was not measured; every cost-per-task figure here is a vendor claim. The article could not verify the task lead's claim of a new Anthropic configurable pricing multiplier for shared pools (it was cut rather than printed). Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks, and OpenAI left the mid-tier Terra slot empty, so this price ladder will move again. Gemini CLI's exact default model name could not be pinned from Google's pricing page (the model headings did not render in the fetch), so the Google column uses the coding-class model rates with that limit disclosed. Pricing pages were checked September 24, 2026; they change.