What does it actually cost to run coding agents full time?

What's a realistic dollar range for running agents full time?

Two real, vendor-published numbers mark the ends of this range, and the gap between them is the whole story.

At the low end: Anthropic's own Claude Code documentation states that across its enterprise deployments, average spend runs about $13 per developer per active day, $150 to $250 per developer per month, and 90% of users stay under $30 a day (code.claude.com/docs/en/costs, checked September 2026). That's a normal engineer running an agent inside their editor on ordinary tickets.

At the high end: OpenAI's own "Research acceleration: The view inside OpenAI" report says that as of mid-August 2026, the median researcher inside its research organization spent more than $600 a day on coding-agent inference at API prices, and the top 10% burned through more than $7,000 a day in tokens (OpenAI, published September 2026; the page blocks automated fetches, so these figures are cross-checked against multiple outlets quoting the report directly). That's a frontier lab running several concurrent, long-horizon agents against its own hardest research problems. Agent runtime there overtook total human labor sometime before June 2026, and by mid-August had climbed to 3.1 agent-workdays of effort for every workday of human labor.

At maybe worth building we look for the number under the headline, and this is it: both groups have access to the same class of frontier models. Nobody at OpenAI is stuck on a worse model than the median Claude Code user. The 46x gap between $13 and $600 a day isn't a pricing gap. It's a usage-pattern gap, and it tells you where your own number will land: closer to $13 if agents handle discrete tickets, closer to $600 if you run several long-horizon agents at once and let them iterate.

A serious agent user, someone who has moved past autocomplete and treats an agent as a coworker for most of the workday, sits between those two numbers. As of September 2026, $30 to $150 a day is the honest range.

What do the actual per-token prices look like right now?

Model prices move fast and most of them are promotional, so treat this table as a snapshot, not gospel. Every row was checked against the vendor's own pricing page or model docs in early September 2026.

Model	Input $/M tokens	Output $/M tokens	Cached input $/M tokens	Source
GLM-5.3-Flash (Z.ai)	$0.075 (promo) list $0.15	$0.25 (promo) list $0.50	$0.015 (promo) list $0.03	docs.z.ai pricing; promo ends Sept 9, 2026
GPT-5.6 Sol (OpenAI)	$4	$20	—	OpenAI model docs; rate holds through Nov 21, 2026
Claude Fable 5.1 (Anthropic)	$10	$50	$0.25 (down 75% from $1)	anthropic.com, Sept 1, 2026
GPT-6 Astra (OpenAI)	$10	$50	$1.00 (90% off)	OpenAI / OpenRouter; 2.5x Sol's rate

The GLM-5.3-Flash promo above expires the same week this page went up. Even at its list price, $0.15/$0.50, it still runs roughly 25 to 60x cheaper per token than either frontier model above, which is the number that actually matters once the promo is gone.

Does the cheapest model per token actually cut your bill?

Not by nearly as much as the token prices suggest, because per-token price and per-completed-task cost are different numbers.

Model	Cost per completed agentic task	Note
GLM-5.3-Flash	~$0.09	Open-weights, MIT license
GPT-5.6 Sol	~$0.95	OpenAI's mid-tier reasoning model
Claude Fable 5	~$3.14	Pre-5.1 pricing; the 5.1 cache-cost cut below narrows this

Figures via Artificial Analysis's Coding Agent Index, which scores full agent stacks (model plus harness) on real software-engineering tasks and reports the average pay-per-token cost to finish one, as analyzed by Matthew Berman in early September 2026. Worth reading with the same skepticism you'd bring to any single-source benchmark: see our take on trusting AI benchmark leaderboards before you make a model swap off one chart.

Read together, GLM-5.3-Flash finishes a task for about 9 cents against Fable 5's $3.14, for a model that scores only 5 to 7 percentage points lower on the underlying intelligence index. That's a 35x cost gap for a single-digit capability gap, and it's the strongest evidence on this page that the sticker price on a model card isn't the number to optimize.

Anthropic's own Fable 5.1 release makes the same point from the other direction. The company says the model is "up to approximately 45%" cheaper for highly agentic work than Fable 5, not because the per-token rate changed (input and output prices are identical at $10/$50) but because cache-read pricing dropped 75%, and agentic work is "context-heavy, tool-heavy work, where cache reads make up most of the cost" (anthropic.com/claude-fable-and-mythos-5-1, Sept 1, 2026). Same headline price. A materially different bill, because agentic workloads don't spend tokens the way a chat session does.

GPT-6 Astra runs the comparison in reverse. It costs 2.5x more per token than GPT-5.6 Sol, and Latent Space's own math puts a fully automated engineer running on Astra at under $6 an hour, cheaper in practice than the lower-priced Sol on comparable work, because it needs less scaffolding and fewer retries to land a correct answer (latent.space/p/astra). The model that costs more per token is not the model that costs more per finished task, which is the entire case for reading a model comparison past the rate card.

So what actually drives the spread?

Three things move your bill, and none of them is which model you picked.

Task length. OpenAI's own report says the mix of work researchers delegate to agents keeps shifting toward "higher level and longer-horizon tasks" over time, and that's a structural reason spend climbs even as per-token prices fall: a longer task burns more tokens, cache misses included, no matter which model runs it.

Concurrency. The same report tracks how many researchers run four or more agents at once, and that number keeps rising. Anthropic's own documentation states plainly that agent teams "use approximately 7x more tokens than standard sessions when teammates run in plan mode," purely because each teammate keeps its own context window. That's a configuration choice, not a model choice, and it's a 7x multiplier before anyone has touched the model picker.

Retry rate. Artificial Analysis has found that the cost of finishing the exact same benchmark task can range from $0.07 to $2.26 with nearly identical output quality, driven by harness and infrastructure choices rather than the underlying model. A bad harness that retries, re-reads files, or fails to hold a warm cache can cost you more than switching from the cheapest model to the most expensive one would have.

Put together: two engineers on the identical model and the identical plan can land 10x apart on their daily bill, purely from how long they let a task run and how many times they let it retry before stepping in.

What should a team actually budget?

Start at $50 to $100 per engineer per day, roughly $1,000 to $2,000 a month per seat, for a team that has moved past pilot mode into agents as a real part of how code ships. That number sits comfortably inside the $30-to-$150 range above and leaves room before you're anywhere near OpenAI's own $600 median.

Don't set the budget by picking the cheapest model on the pricing page. Set it by watching concurrency and task length, the two levers that actually move the number, and instrument spend before you scale the headcount running on it. That's the same advice Anthropic gives its own enterprise customers: pilot with a small group, track it, then roll out wider once you know your own baseline.

Treat crossing roughly $300 a day per engineer as a signal to check how long tasks are running and how many agents are running at once, not as a signal to swap models. If the number keeps climbing anyway, that's not automatically a red flag. OpenAI's own research organization runs its median researcher at $600 a day and its top 10% past $7,000, and reports that experiments per active researcher hit an all-time high in August 2026 because of it.

Frequently asked questions

How much does it cost to run a coding agent like Claude Code all day, every day?

Anthropic's own enterprise data puts the average at about $13 per developer per active day and $150 to $250 per developer per month, with 90% of users staying under $30 a day. That's typical usage inside an editor on normal tickets, not a team running several long-horizon agents at once.

Why does OpenAI's own research org spend so much more per day than a typical Claude Code user?

Because it isn't typical usage. OpenAI's median researcher spends more than $600 a day and the top 10% spend more than $7,000, running several concurrent agents on long-horizon research tasks instead of discrete tickets. Same class of frontier models as everyone else, far higher task length and concurrency.

Does switching to a cheaper model, like GLM-5.3-Flash, actually cut your agent bill?

It helps, but less than the per-token price implies. GLM-5.3-Flash finishes an agentic task for about 9 cents against roughly $3.14 for Claude Fable 5 on Artificial Analysis's Coding Agent Index, a real gap, but harness and task-length choices can swing a single model's cost on the same task by more than that.

Is GPT-6 Astra worth paying 2.5x more per token than GPT-5.6 Sol?

Often yes, on a per-finished-task basis. Latent Space's own calculation puts an automated engineer running on Astra at under $6 an hour, because it needs fewer retries and less scaffolding to land a correct answer. The higher rate card doesn't translate into a higher total bill.

What's the single biggest lever for controlling coding-agent spend?

Task length and retry rate, not model choice. Anthropic's own documentation says agent teams use about 7x more tokens than a standard session when teammates run in plan mode, purely from running more concurrent context windows. That's a configuration choice, and it dwarfs any model swap.

How much should a team budget per engineer per month for coding agents?

$1,000 to $2,000 per engineer per month ($50 to $100 a day) is a reasonable working default for a team using agents as a real part of the workflow in late 2026. Pilot with a small group first and set the real number from your own measured baseline, not a guess.

Will agent costs come down as per-token model prices keep falling?

Per-token prices are falling fast, GLM-5.3-Flash lists at a fraction of a cent, but total spend has been rising anyway, because task length and concurrency are growing faster than price is falling. OpenAI's own agent runtime passed total human labor inside its research org before June 2026 and kept climbing after.