The coding agent bill is the harness, not the model
Sourcegraph cut credit prices, Baseten cut tool-loading tokens, and Toloka showed the same model costing 4.4× more under a different harness. Teams should price the harness first.
- Sourcegraph lowered Deep Search to 50 credits per search and Code Finder to 3 credits per call, and Smart Hovers will no longer cost credits. Self-reported m:…2639
- Baseten says deferred tool loading makes coding agents up to 62.5% cheaper, because eager loading with 3 MCP servers added about 141k tokens to the system prompt while deferred loading cut tool schema tokens to about 1.3k. Self-reported m:…2889
- Toloka found the same model on the same tasks cost $20 under Toloka Forge and $82 under Harbor, and harness choice changed the cost for every model tested by up to 4.4×. Self-reported m:…7584
- In a preprint, SuperScout matches the best single model solve rate on SWE-bench Pro, 159 of 266 tasks, at about a fifth of the total cost per solve. Self-reported m:…0961
What changed
Sourcegraph is cutting what its coding tools cost in credits. Deep Search drops to 50 credits per search, Code Finder drops to 3 credits per call, and Smart Hovers will no longer cost credits. Sourcegraph says the lower costs come from improved efficiency in Deep Search and newer, lower-cost models. Service account billing follows the same model as human users starting November 19, 2026. m:…2639
Baseten makes a similar point from the other side of the bill. Its deferred tool loading passes tool names plus a tool_search tool first, instead of injecting the full list of tool schemas into the system prompt. Baseten says the change makes coding agents up to 62.5% cheaper and up to 26.2% faster, with agentic benchmarks saving an average of 51.1% in costs and 14.9% in time. The reason is size: eager loading with 3 MCP servers added about 141k tokens to the system prompt, while deferred loading cut tool schema tokens to about 1.3k. m:…2889
The harness sets the price
Toloka tested what happens when the model stays the same and only the harness changes. Running Haiku 5.5 on 50 Terminal-Bench tasks cost $20 under Toloka Forge and $82 under Harbor, and harness choice changed the cost for every model tested by up to 4.4×. Harbor trials needed 2.7× more machine time than Toloka Forge trials across the board. Accuracy moved too: three of ten models showed accuracy differences of 13 to 18 points between harnesses, though for 7 of 10 models the accuracy differences were not measurable. m:…7584
Two preprints reach the same conclusion from the research side. In a preprint, a routing design called SuperScout matches the best single model solve rate on SWE-bench Pro, 159 of 266 tasks, at about a fifth of the total cost per solve, with the searcher adding less than half a cent of GPU time per task. In another preprint, harness swaps change task outcomes at a rate comparable to simple reruns, and inference cost differences between harnesses range from under 10% to 3-4x depending on the model. The implication is that the harness is not overhead on top of the model bill. In many setups it is the bill. m:…0961 m:…8345
The savings are not automatic. OpenAI and Anthropic APIs typically support deferred tool loading, but Baseten notes Claude Code turns off deferred tool loading when using third-party open model providers. Toloka Forge was cheaper than Harbor on 8 of 10 models tested, and Harbor held a cost advantage on GPT-6 Astra, driven by a 91% cache hit rate compared with 84% on Toloka Forge. The harness that wins depends on the model, the cache, and the workload, which is why the comparison has to be run rather than assumed. m:…2889 m:…7584
What to budget instead
The practical read is that teams should price the harness before they price the model. Sourcegraph and Baseten both say their savings come from doing less work per call, not from a cheaper model, and Toloka shows the same task can cost up to 4.4× as much under a different harness. That also means benchmark scores need the harness named next to them, since Harbor reported 74 trials that hit the turn cap as normal finishes, making them indistinguishable from successful trials. A score without its harness is a price without its receipt. m:…2639 m:…7584
None of this says the model does not matter. It says the model is priced through a layer most budgets never see. The teams that will spend least are the ones that measure tokens per task under their own harness, compare harnesses the way they compare models, and treat a credit-price cut as a prompt to re-measure rather than a reason to stop measuring. m:…7584 m:…2889
What is still unresolved
- Whether Sourcegraph's lower credit costs hold once the newer, lower-cost models behind them change.
- How Toloka's harness cost gap looks on tasks outside the 50 Terminal-Bench tasks tested.
- Whether deferred tool loading helps as much for agents with small tool sets as for agents with 3 MCP servers attached.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.