OpenCode removes two DeepSeek routes linked to upstream filter errors

• OpenCode removed two DeepSeek routes whose upstream content filters were producing excessive errors, according to thdxr's follow-up.
• Benchmark results were broadly similar across configurations, although OpenCode's table contains two sharp 100K-context outliers.
• OpenCode's own GPU cluster remains unstable and capacity constrained, so some requests may still reach DeepSeek's official API, thdxr said.
• A $10 Go subscription covered a roughly 1.2-billion-token coding task that teortaxesTex described as a slog.

Go asks clients for a stable session header specifically to optimize routing and prompt caching. OpenCode also kept DeepSeek's opt-in while its own GPU cluster remained unable to handle all traffic.

OpenCode's thdxr attributed the two problematic configurations to upstream content filters that were firing too often, and said both configurations had been removed. Affected users received errors.

The earlier announcement said the configurations would be retired; this follow-up reported the removal. Neither post attributed the broader complaints about weaker answers to those filters.

OpenCode compared seven configurations of DeepSeek V4.1 Flash against the official DeepSeek API with thinking enabled, and included a genuinely useful extra column: traffic share.

Three comparisons stand out:

• Overall accuracy: OpenCode configurations ranged from 89.3% to 90.8%, versus 90.5% for the official API.
• Busiest route: Config 1 carried 58% of traffic and scored 90.8% overall, 0.3 percentage points above the official API.
• 100K-context outliers: Configs 6 and 7 scored 66.7% and 73.3%, respectively, versus 100% elsewhere. Together, those two rows accounted for 12% of traffic, not a measured 12% error rate.

Testing the model/provider combination is part of OpenCode's published selection process, which explicitly acknowledges that provider configurations can change performance and quality.

The chart's tool scores cluster between 99.3% and 99.7%. It labels the probes simply “Tools” and “100k ctx,” without task descriptions, sample sizes, or a multi-turn coding-agent evaluation.

Long-context use and tool calling were likelier places for DeepSeek to fail than one-shot completions, teortaxesTex argued.

The broader user reports were “probably real,” according to thdxr's follow-up, but OpenCode still needed specific examples to investigate them.

Config 1 was OpenCode's own GPU cluster, according to thdxr, who said the numbered configurations mostly represented request destinations.

OpenCode attributed slower token generation to limited hardware capacity. The cluster also remained unstable and unable to handle peak demand, leaving some requests eligible for routing to DeepSeek's official API.

Thdxr said he was hesitant to remove the opt-in while that routing remained possible. He also said more GPUs were needed to bring throughput up to DeepSeek's level.

A separate coding task involving Ballance consumed roughly 1.2 billion tokens at an estimated $11.62 in API-equivalent charges. In teortaxesTex's account, the $10 Go subscription made the run negligibly cheap, but progress remained a slog.

More than 99% of the recorded tokens were cache-hit input, calculated from the dashboard's breakdown. The author judged that Claude or GPT would have been faster and possibly cheaper.

Thdxr asked users to DM three fields:

• What happened.
• When it happened.
• The organization ID, available in the console.

Similar reports have helped OpenCode find real issues before. Separately, the Go integration documentation specifies a stable x-opencode-session value for each conversation so the service can optimize routing and prompt caching.