OpenCode Go revises limits after DeepSeek price increase

• OpenCode Go revised its limits after DeepSeek raised prices, according to OpenCode's announcement.
• Because Go defines allowances in dollars rather than as a fixed request pool, the change OpenCode announced moves the number of supported requests with model cost.
• DeepSeek's V4 rates now vary by time of day, at half price off-peak, according to DeepSeek's API announcement; thdxr's hosting update says getting back near the former economics is difficult.
• Cache reads are the cost center thdxr identified in a reply, a consequential detail for coding-agent sessions with long reused context.
• OpenCode has started what it calls Operation Cheepseek, while thdxr says several hosting configurations are under test.

OpenCode's DeepSeek data page reports V4 Flash as 73% of observed Go volume last week. The Go documentation serves V4 Flash and V4 Pro through OpenAI-compatible Chat Completions endpoints.

Go costs $5 for the first month and $10 per month afterward. Its usage-limit documentation sets a $12 five-hour ceiling, $30 weekly ceiling, and $60 monthly ceiling, then states that request counts vary by the selected model's cost.

The same table estimates 3,800 V4 Flash requests per five-hour window, 9,450 per week, and 18,900 per month under its observed request mix. The V4 Pro estimates are 1,050, 2,600, and 5,200 respectively, per OpenCode's limit table.

DeepSeek's API announcement made the new V4 pricing effective August 17. A TechNode pricing report puts peak time at 01:00 to 04:00 and 06:00 to 10:00 UTC, or 09:00 to noon and 14:00 to 18:00 Beijing time, with all other hours billed at half the peak rate.

OpenCode's cost table lists these per-million-token rates in input, output, cached-read order:

• V4 Flash: off-peak $0.22, $0.66, $0.007. Peak $0.44, $1.32, $0.014.
• V4 Pro: off-peak $0.66, $1.98, $0.022. Peak $1.32, $3.96, $0.044.

OpenCode's observed-session assumptions put a typical V4 Flash request at 410 uncached input tokens, 71,300 cached input tokens, and 310 output tokens. V4 Pro is modeled at 750 uncached, 82,000 cached, and 290 output.

That makes cached input roughly 174 times the uncached input volume for Flash and 109 times for Pro. The provider's cache-read rate therefore sits inside the dominant input category for the request shapes OpenCode uses to calculate its limits.

OpenCode called the effort Operation Cheepseek in its update. thdxr said the team was testing several setups to host DeepSeek near the previous price, and that providers claiming they had already done so were not actually matching it.

The hosting economics were “not impossible,” thdxr wrote, but required a specific hardware layout and physical infrastructure. Neither OpenCode's post nor thdxr's test update identifies a provider or configuration.