Cheapest way to run Claude Code: Requesty's benchmark

As of November 2026, the cheapest way to run Claude Code is to start with Requesty's allowance of 200 requests per day on free models, available through its Claude Code gateway integration. Among paid models in Requesty's coding benchmark, GPT-6 Luna has the lowest cost per solve at $0.07. Qwen3.8 Flash follows at $0.08, then Claude Haiku 5.5 at $0.14. The paid ranking comes from our benchmark agent, so treat it as a starting point for your own Claude Code tests.

Models ranked by cost per solve

For our coding benchmark, we put 25 models through 129 private tasks from Requesty's Go backend. Each model used the same agent and tools (bash, read, edit and write), with an 80-turn budget. Both the hidden tests and existing tests had to pass before we counted a task as solved.

The November 2026 table below selects models from the ranking and sorts them by cost per solve, using the benchmark updated on October 29. The benchmark uses list prices from the cheapest Requesty provider for each model, and the run cost includes retries.

The cheapest model leaves more work unfinished. GPT-6 Luna solved 65.1% of the benchmark tasks, while Haiku solved 81.4% for an additional $0.07 per solve. When Luna cannot finish a task, a successful run with another model adds to the cost.

What 100 solved tasks could cost

Take 100 solved tasks: multiply the November 2026 cost per solve by 100. This extrapolates from the benchmark workload, and failed runs are already included in the cost per solve.

Opus costs about seven times as much as Haiku per solve, while solving 10.8 percentage points more tasks. Whether that extra spend earns its keep depends on the tasks you give it and how you handle the ones Haiku leaves unfinished.

These estimates use the benchmark's provider prices. Requesty charges the provider's token price plus 5%, so a $14 token bill becomes $14.70 through Requesty. See Requesty's Claude Code pricing.

List price leaves out the work

The listed input and output token prices cover only part of a coding bill. A model can take more turns, generate more output or hit the cache less often before finishing a task.

Source: Requesty coding benchmark, November 2026 prices.

GPT-6 Sol and GPT-6.1 Sol have identical listed token prices, but the newer model costs less than half as much per solve on this workload. Mistral Large 4 has lower listed input and output prices than either Sol model, yet costs $3.04 per solve. It had 0% cache hits and averaged 46.9 turns per task in the benchmark.

Output volume can pile up too. MiniMax M3.1 Flash generated 195.9k output tokens per task across 65.5 turns, despite a listed output price of $0.60 per million. We break down the calculation behind cost per task and cost per token separately.

Hard tasks change the model choice

The benchmark includes 30 extra hard tasks and 18 extreme tasks. Opus leads both groups, while the cheapest model in the overall ranking solves a smaller share.

Source: Requesty coding benchmark.

On easy and medium work, a lower-cost model is a sensible starting point. On the hardest tasks, Opus trades some cost efficiency for a higher solve rate. Test whether that pattern holds on your own repository.

How cache hit rates change the input bill

Prompt caching lets a provider reuse prompt content it has already processed. That can lower the price of repeated input in coding sessions, where an agent sends context across multiple turns.

In Requesty's October 2026 data, Cline had a 96.00% cache hit rate and Claude Code had an 85.22% hit rate. Agents were identified by user agent and app title; each included agent had at least 1,000 calls and five organizations.

What does that gap do to an input bill? Take $2 per million fresh input tokens, cached reads at 10% of the fresh price and no cache-write charge.

Effective input price = fresh price x ((hit rate x 0.1) + uncached share)

At these prices, the Claude Code hit rate produces an input bill about 71% higher than the Cline hit rate. Output tokens and cache writes would add to the bill.

The benchmark has a separate cache result: Sonnet 5.5 had 99.2% cache hits with our benchmark agent. Your Claude Code sessions can look different. For the mechanics behind the calculation, see prompt caching versus semantic caching.

Subscription or API?

With a subscription, you pay a monthly plan price. With an API or gateway, token usage drives the bill. At November 2026 prices, Claude's pricing page lists Pro at $20 per month billed monthly, Max at $100 and Ultra at $200.

API billing tends to suit occasional use and automated jobs, while a subscription tends to suit daily use. Requesty also lists an allowance of 200 requests per day on free models. Claude Code can connect to Requesty through its gateway integration.

For a usage-based spending reference, Anthropic reports about $13 per developer per active day and $150 to $250 per developer per month across enterprise deployments. Separately, Requesty's October 2026 data puts Claude Code users at an average of $131 per month; in that dataset, one API routing key counts as one user.

Those subscription prices and usage averages describe different groups and billing arrangements. Compare your own usage with the plan price before choosing.

Run Claude Code with a cheaper model

Requesty's Claude Code guide covers the gateway setup. The example below uses GPT-6.1 Sol, which solved 86.8% of benchmark tasks at $0.30 per solve using November 2026 prices.

1. 1
   Create a Requesty key

   Create a Requesty API key and store it in the REQUESTY_API_KEY environment variable. You can set a monthly spend cap for the key.

2. 2
   Point Claude Code at the gateway

   Set ANTHROPIC_BASE_URL to the Claude Code gateway endpoint at https://router.requesty.ai. Set ANTHROPIC_AUTH_TOKEN to the value of REQUESTY_API_KEY.

3. 3
   Select the model

   Set ANTHROPIC_MODEL to openai/gpt-6.1-sol, then start Claude Code. Requesty's guide recommends the environment variable because Claude Code can reject model names it does not recognize in the /model command.

4. 4
   Measure your own work

   Run a sample of tasks from your repository and record which ones pass your tests. Use Requesty's Logs to inspect per-request cost and cache hit rates, then divide the total spend by the number of tasks solved.

Which model should you try first?

Set a spend cap on the gateway key, then run a sample of tasks from your own repository. Compare the total bill with the number that pass your tests, and track the cache hit rate alongside both.