GPT-6.1 Sol Lands on Devin: Same List Price, 95% Drop in Cache Read Costs

GPT-6.1 Sol Is Now Available in Devin

OpenAI has released GPT-6.1 Sol, which is now live in both Devin Desktop and Devin CLI. For developers relying on Devin for AI-assisted coding, this is a noteworthy update — not just because of the model itself, but because of the significant shift in its underlying cost structure.

On the pricing front, GPT-6.1 Sol maintains the exact same list price as its predecessor, GPT-6 Sol: $2 per million input tokens and $10 per million output tokens. The sticker price hasn't changed — but the real difference is buried in the caching mechanics.

What a 95% Drop in Cache Read Costs Actually Means

The most substantive change in this update is a 95% reduction in cache read costs. For coding workloads in particular, this is a big deal.

In real-world coding scenarios, models frequently re-read the same codebase, context files, or system prompts across multiple turns. When those repeated reads hit the cache, they can be reused at a fraction of the original cost. A 95% price cut on cache reads means that tasks heavily reliant on long contexts and repeated code references will see their actual spend drop dramatically.

For autonomous coding agents like Devin, a single task typically involves multiple rounds of reasoning and repeated context lookups. The higher the cache hit rate, the more money saved. This explains how per-task costs can fall significantly even when the list price stays the same.

What is prompt caching? Prompt caching is an infrastructure-level optimization for LLM APIs. When consecutive requests share the same prefix content — such as system prompts, codebase summaries, or long documents — the server can retain the key-value computation results (KV Cache) in VRAM or fast storage and reuse them for subsequent requests instead of recomputing. From the user's perspective, tokens that hit the cache are billed at the "cache read" rate, which is typically far lower than the standard input price. OpenAI, Anthropic, and other major providers all offer similar mechanisms. In AI coding agent scenarios, system prompts and repository context can span tens of thousands to hundreds of thousands of tokens and remain largely unchanged across multiple calls within a single task — making cache hit rates naturally high and the savings most pronounced.

FrontierCode 1.1 Performance: Same Score, Half the Cost

GPT-6.1 Sol's results on the FrontierCode 1.1 coding benchmark are compelling. According to official data, it matches GPT-6 Sol's scores across all reasoning effort levels while costing roughly half as much per task.

In other words, users don't have to trade quality for cost — they get equivalent results at half the price.

The numbers at the low reasoning effort tier are especially striking. At the low effort setting, GPT-6.1 Sol achieves a score of 58.1% at just $0.21 per task — making it the highest-scoring model among all those priced under $0.30 per task. For teams running high-throughput, cost-sensitive batch coding workloads, this price-performance sweet spot is highly attractive.

What is FrontierCode 1.1? FrontierCode 1.1 is a coding benchmark designed to evaluate code generation capabilities across real-world software engineering tasks, including function completion, bug fixing, and multi-file editing. Compared to earlier benchmarks like HumanEval and SWE-bench, it more closely reflects the complexity of production environments. Reasoning effort is a parameter offered by some model APIs that lets callers trade off reasoning depth against response speed and cost. High effort triggers more internal chain-of-thought steps; low effort generates answers faster and more lightly. This parameter is especially useful for automated batch tasks — not every coding subtask requires maximum reasoning depth, and dialing it down can significantly reduce per-call costs without meaningfully impacting quality.

What This Means for Developers

Taken together, the logic behind GPT-6.1 Sol's update is straightforward: rather than attracting users with a lower list price, it reduces real-world costs by optimizing cache economics and inference efficiency. This approach benefits heavy users most — particularly teams running long-running autonomous coding tasks.

For teams already using Devin, switching to GPT-6.1 Sol is essentially a free upgrade — same interface, same pricing structure, but with close to half the cost in real workflows. For teams evaluating AI coding tools, the "same score, half the cost" data point on FrontierCode 1.1 offers a relatively objective benchmark for comparison.

One caveat: this information is primarily based on official release materials, and the specific methodology behind FrontierCode 1.1 and the details of the comparison models have not been fully disclosed. Developers making technology decisions are still advised to validate results against their own codebases and actual task distributions.