Token usage

Experiment totals

• Coverage: each row contains 150 runs. Total tokens are input plus output.
• OpenAI costs: API-equivalent estimates, not invoice totals. Rates per 1M uncached input / cached input / output tokens are GPT-5.5 $12.50 / $1.25 / $75, Sol $10 / $1 / $60, and Luna $2 / $0.20 / $12.
• DeepSeek price change: the old rate was $0.435 / $0.003625 / $0.87. Since August 16, peak is $1.32 / $0.044 / $3.96 and off-peak is $0.66 / $0.022 / $1.98. Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday-Friday.
• Two DeepSeek costs: “Estimated actual cost” uses the price in force during each run; “Same tokens at current DeepSeek rates” reprices the same work under today's peak/off-peak schedule.
• Timing estimate: usage is reported per run, so tokens are spread evenly over runs that crossed a price boundary.
• Why the DeepSeek Galaxy token total is high:
  1. Galaxy tools were provided: the experiment included dedicated Galaxy MCP tools.
  2. The setting came from DeepSeek's official setup: DeepSeek's official Codex setup script included supports_search_tool: true in its model catalog.
  3. A Codex bug then hid the tools: supports_search_tool tells Codex whether a model can search for tools on demand. When it is enabled, Codex does not supply all tools at the start; it waits for the model to call tool_search. DeepSeek did not make that call, so it could not see the Galaxy MCP tools.
  4. 57 of the 150 reported Galaxy runs used a fallback: the agent sent the same operations to the Galaxy API through shell commands. The scientific calculations still ran on usegalaxy.org. These runs averaged 14.55M tokens and 70.2% accuracy; their matched Open-ended code runs averaged 2.57M tokens and 59.6% accuracy.
  5. The fallback increased tokens: command output accumulated in the conversation and was sent back to the model on later steps, repeatedly enlarging the input context.
  6. 93 of the 150 reported Galaxy runs used the normal route: after we disabled supports_search_tool, Codex supplied the Galaxy MCP tools directly. These runs averaged 4.29M tokens and 89.2% accuracy; their matched Open-ended code runs averaged 2.02M tokens and 93.5% accuracy.
  7. Overall effect: DeepSeek used 1.23B tokens with Galaxy versus 334.9M in Open-ended code, about 3.7 times as many. Of its 1.223B Galaxy input tokens, 1.207B were cache hits (98.7%), showing that the increase mainly came from repeated context rather than 3.7 times more scientific work or reasoning.

Prices checked August 27, 2026: OpenAI Priority · OpenAI models · DeepSeek pricing.

Item totals

Each cell aggregates the three replicates for one model and execution condition.