AI agent cost comparison: 50-step runs by model

Across my measurements, the five models spread widely on cost per 50-step agent run. DeepSeek V4 Pro 0813 is the cheapest at 30¢, and Kimi K3 is the most expensive at $2.24, even though Kimi has the lowest input price of the five. For agent work, the per-token price on the label tells you little. What matters is how many tokens the model spends across 50 steps, including reasoning tokens, which are billed as output.

How I tested it

The agent scenario is an 8,000-token system prompt, 50 steps, 1,500 tokens of extra context per step, and 300 visible output tokens per step. The 50-step agent numbers are modeled for all five models, so treat them as estimates built from that scenario, not as a log of 50 live calls.

The SQL query task is a smaller single-shot test. GPT-6.1 Sol, Claude Sonnet 5.5 and Grok 4.7 are measured there, while DeepSeek V4 Pro 0813 and Kimi K3 are estimated. Measured runs used default settings, one run per task, with output capped at 16,000 tokens. Details are on the methodology page.

Prices are USD per 1M tokens, as of 2026-10-19.

Model	Input / output / cached input	Context	LMArena Text
GPT-6.1 Sol	$2.00 / $10.00 / $0.10	1,050,000	1484
Claude Sonnet 5.5	$2.00 / $10.00 / $0.10	1,000,000	1475
DeepSeek V4 Pro 0813	$0.66 / $1.98 / $0.022	1,048,576	not listed
Grok 4.7	$2.00 / $6.00 / $0.50	500,000	1444
Kimi K3	$0.25 / $11.94 / $0.21	1,048,576	1487

Cost per 50-step run

The median of all 336 models on this task is 48¢ per run. Here is where the five land:

#	Model	Cost per task	WebDev score	Tokens in / out	Basis
1	DeepSeek V4.1 FlashDeepSeek	2.0¢$19.58 / 1,000	1618	2,235,262 / 102,435	modeled best value
2	GLM 5.3 FlashZ.ai	2.6¢$25.63 / 1,000	1606	2,239,737 / 43,814	modeled
3	GPT-6 LunaOpenAI	4.1¢$40.71 / 1,000	1578	2,237,500 / 22,000	modeled
4	MiMo-V2.6-ProXiaomi	6.5¢$64.51 / 1,000	1625	2,241,975 / 24,380	modeled best value
5	Claude Haiku 5.5Anthropic	6.6¢$66.35 / 1,000	1585	3,101,175 / 50,340	modeled

DeepSeek V4 Pro 0813: 30¢ per run, $303.94 per 1,000 runs.
GPT-6.1 Sol: 53¢ per run, $528.60 per 1,000 runs.
Claude Sonnet 5.5: 84¢ per run, $841.64 per 1,000 runs.
Grok 4.7: $1.70 per run, $1,702.19 per 1,000 runs.
Kimi K3: $2.24 per run, $2,238.34 per 1,000 runs.

Caching matters a lot. Without caching, DeepSeek V4 Pro 0813 would cost $1.68 per run and GPT-6.1 Sol $4.62. Agents resend the same system prompt and history on every step, so cached input is the normal case, not an edge case. Still, check that your provider actually applies the cache to your traffic.

Why Kimi K3 looks cheap and isn't

Kimi K3 has the lowest input price in this group at $0.25 per 1M tokens. Its output price is $11.94, and the modeled run includes 132,350 reasoning tokens, billed as output. The result is $2.24 per run, the most expensive of the five. On the SQL task it is also the most expensive of the five at 3.8¢ per query, estimated. Its LMArena Text score of 1487 is strong, so it is capable. It just spends a lot of tokens to get there.

Why Sonnet 5.5 costs more than GPT-6.1 Sol

The two have identical list prices: $2.00 input, $10.00 output, $0.10 cached input. The difference is tokens. Anthropic's tokenizer ratio is 1.39 against OpenAI's 1.00 for the same English text, so the Sonnet run used 3,101,175 input tokens against 2,237,500 for Sol. Sonnet also used 10,900 reasoning tokens, while the modeled Sol run used 0. Same price on paper, 84¢ versus 53¢ per run.

Check the tokenizer and the reasoning mode, not only the price sheet.

Grok 4.7 and the SQL task

Grok 4.7 costs $1.70 per agent run. Its output price of $6.00 is lower than Sonnet's, but its cached input price is $0.50 against $0.10, and the run used 2,403,075 input tokens and 45,450 reasoning tokens. Its LMArena Text score of 1444 is the lowest of the four models here that have one.

On the SQL query, the measured costs were:

Model	SQL cost per query	Basis
DeepSeek V4 Pro 0813	0.43¢	estimated
GPT-6.1 Sol	0.48¢	measured
Claude Sonnet 5.5	1.1¢	measured
Grok 4.7	1.2¢	measured
Kimi K3	3.8¢	estimated

For agent work, I would not pick Grok 4.7 over Sol on cost.

DeepSeek V4 Pro 0813: cheapest, with caveats

DeepSeek V4 Pro 0813 was cheapest on both tasks: 30¢ per agent run and 0.43¢ per SQL query. It used 87,450 reasoning tokens in the agent run, so the low cost comes from the rate, not from thinking less. But the numbers are modeled or estimated, and the no-cache figure of $1.68 shows how much it depends on caching. Test on your own workload before committing.

What subscriptions cover

Subscriptions do not replace API billing for these runs. Plans relevant to coding-agent use:

ChatGPT Plus: $20 per month, with Codex on the web, in the CLI, in the IDE extension, and on iOS. API usage is billed separately.
ChatGPT Pro 100: $100 per month, including Pro models and Codex.
Claude Pro: $20 per month, with Claude Code included.
Claude Max 5x: $100 per month, with $100 per month in API credits for the Claude Platform.
Claude Max 20x: $200 per month, with $200 per month in API credits for the Claude Platform.
Cursor Pro: $20 per month, with extended limits on Agent.
GitHub Copilot Pro: $10 per month, with $15 per month total GitHub AI Credits.

I found no subscription plans for Kimi K3 or DeepSeek V4 Pro 0813 in the plan data, so those are pay-per-token through an API. For steady coding work, a flat plan may be easier to budget. See plans and the calculator to model your own volume.

Verdict

Worth the cost for agent work: GPT-6.1 Sol. It costs 53¢ per modeled run and 0.48¢ per measured SQL query, with an LMArena Text score of 1484.
Cheapest per run: DeepSeek V4 Pro 0813 at 30¢. Verify it on your own tasks, since the agent figure is modeled.
Looks cheap per token but isn't: Kimi K3. The $0.25 input price misleads, because the $11.94 output price and heavy reasoning push a run to $2.24.
Costly for what it adds here: Claude Sonnet 5.5 at 84¢ per run, because of tokenizer and reasoning overhead. Grok 4.7 at $1.70 is harder to justify.

See the agent task page and the SQL query page for more models, and news for price changes.

FAQ

What does a 50-step agent run cost?

In my modeled scenario: 30¢ with DeepSeek V4 Pro 0813, 53¢ with GPT-6.1 Sol, 84¢ with Claude Sonnet 5.5, $1.70 with Grok 4.7 and $2.24 with Kimi K3. The median across 336 models was 48¢.

Why does a cheap input price not mean a cheap agent?

Agents also generate output and reasoning tokens over many steps. Kimi K3 has the lowest input price here, but its $11.94 output price and 132,350 reasoning tokens made its run the most expensive of the five.

Do subscription plans cover API costs for agents?

Not in general. ChatGPT Plus says API usage is billed separately. Claude Max 5x and 20x include $100 or $200 per month in API credits for the Claude Platform, but the runs here were priced per token.