As of November 2026, DeepSeek V4.1 Flash is our first model to test for a backend coding agent: it solved 77% of tasks at $0.41 per solve. GLM-5.3 Flash cost less at $0.38 per solve but solved 64%, while Kimi K3.1 Pro led all 11 candidates with an 81% success rate at $1.04 per solve. The figures come from our benchmark of 129 private Go backend tasks.
Which models cost least per solve?
Cost per task includes retries, while cost per solve accounts for the share of tasks that passed the tests. The ranking covers candidates with sourced open-weight releases, using published results and November 2026 prices from the Requesty coding benchmark.
| Rank | Model | Tasks solved | Cost per task | Cost per solve |
|---|---|---|---|---|
| 1 | GLM-5.3 Flash | 64% | $0.24 | $0.38 |
| 2 | DeepSeek V4.1 Flash | 77% | $0.31 | $0.41 |
| 3 | Step 5 Preview | 57% | $0.28 | $0.48 |
| 4 | Hy4 Preview | 74% | $0.38 | $0.51 |
| 5 | MiniMax M3.1 Flash | 67% | $0.37 | $0.55 |
| 6 | Muse Spark Open | 71% | $0.61 | $0.87 |
| 7 | Kimi K3.1 Pro | 81% | $0.85 | $1.04 |
| 8 | GLM-5.3 | 67% | $1.46 | $2.16 |
| 9 | Kimi K3 | 72% | $1.72 | $2.39 |
| 10 | DeepSeek V4 Pro | 68% | $1.82 | $2.66 |
| 11 | Mistral Large 4 | 78% | $2.36 | $3.04 |
A lower cost per solve can come with a lower success rate. GLM-5.3 Flash is cheapest here, but DeepSeek V4.1 Flash solved 13 percentage points more tasks for $0.03 more per solve. The right tradeoff depends on what a failed task costs you, including whether someone has to finish it by hand.
What did the benchmark measure?
We put 25 models to work on changes to our Go backend. Every run used the same coding agent, with bash, read, edit and write tools, and an 80-turn budget. A task counted as solved when its hidden tests and existing tests passed, according to the benchmark methodology.
The task set contained 14 easy, 28 medium, 39 hard, 30 extra hard and 18 extreme tasks. Token prices were list prices from the cheapest Requesty provider for each model. We measured timing through Requesty in one region, so the figures include provider queueing and network time.
These results measure backend coding tasks in one agent setup. They are a starting point for evaluating coding agents, including their use of tools, but a different workflow needs its own tests.
Read the full Requesty coding benchmark for the task design and complete model comparison. For a comparison that also includes proprietary models, see best AI coding models by benchmark and cost.
Which model should you try first?
Start with DeepSeek V4.1 Flash for its 77% success rate at $0.41 per solve. Try GLM-5.3 Flash if the lowest cost per solve is the main constraint: it solved 64% at $0.38 per solve. For the highest success rate among these candidates, test Kimi K3.1 Pro, which solved 81% at $1.04 per solve.
Difficulty changes the picture. In our benchmark, Kimi K3.1 Pro solved 100% of hard tasks and 39% of extreme tasks. Mistral Large 4 solved 95% and 22%, while DeepSeek V4.1 Flash solved 97% and 17%.
A team whose backlog resembles the extreme tasks may value those extra solves more than a lower average cost. Test that against your own task mix: the benchmark includes 18 extreme tasks.
Why do token prices and cost per solve differ?
An agent can make multiple model calls, read files and revise code before a task ends. The price of a million input or output tokens is one part of the bill; the number of tokens used and the share of tasks that pass tests also affect cost per solve.
These November 2026 benchmark prices are per million tokens, using the cheapest Requesty provider listed for each model.
| Model | Input price | Output price | Cost per task | Cost per solve |
|---|---|---|---|---|
| GLM-5.3 Flash | $0.12 | $0.46 | $0.24 | $0.38 |
| DeepSeek V4.1 Flash | $0.15 | $0.60 | $0.31 | $0.41 |
| Kimi K3.1 Pro | $2.00 | $10.00 | $0.85 | $1.04 |
| Mistral Large 4 | $0.68 | $2.09 | $2.36 | $3.04 |
Take Mistral Large 4 and Kimi K3.1 Pro: Mistral's listed input and output prices were both lower, yet its cost per solve was $3.04 against Kimi's $1.04. Compare the cost of completed work alongside the price list.
What would 100 solves cost?
At the benchmark's cost-per-solve rates, 100 successful tasks would cost 100 × $0.41 = $41 with DeepSeek V4.1 Flash and 100 × $3.04 = $304 with Mistral Large 4. That is a $263 difference.
Your tasks and agent will produce different numbers, so recalculate with your own spend and verified outcomes.
Provider choice can also change the economics when the weights are the same. Our provider variance analysis covers that comparison.
How do speed and caching affect the choice?
Full task time includes more than the time spent waiting for an individual model turn. In the benchmark, GLM-5.3 Flash averaged 1.0 second per turn and 11.4 minutes per task. DeepSeek V4.1 Flash averaged 2.3 seconds per turn and 16.5 minutes per task.
| Model | Average time per task | Benchmark cache hit rate |
|---|---|---|
| DeepSeek V4.1 Flash | 16.5 minutes | 97% |
| Mistral Large 4 | 10.6 minutes | 0% |
The agent also takes a different number of turns with each model. Kimi K3.1 Pro averaged 37 turns and 2.9 million tokens per task; Mistral Large 4 averaged 47 turns and 3.4 million tokens; DeepSeek V4.1 Flash averaged 45 turns and 3.6 million tokens, according to the same benchmark.
Cache results depend on the setup. In our April 2026 coding-agent data, prompt-cache hit rates ranged from 45.5% for Kilo Code to 91.9% for Claude Code. The agent architecture and model provider both affect the rate, so measure it with the agent you plan to run.
What does production traffic add?
Our production rankings show a separate view of DeepSeek V4.1 Flash and GLM-5.3 Flash from August 31 through November 2, 2026.
| Model | Token share | Spend share | Cache hit rate | Median streaming time to first token | Server error rate | Blended cost per million tokens |
|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | 14.1% | 2.8% | 88.2% | 1,170 ms | 1.8% | $0.09 |
| GLM-5.3 Flash | 8.6% | 1.4% | 90.9% | 1,218 ms | 0.2% | $0.07 |
The blended costs are charged rates that include prompt-cache discounts and the token mix.
What should you check before deploying?
License terms differ across open-weight releases. The GLM-5.3 Flash repository publishes an MIT license. The Mistral Large 4 model card specifies the Mistral Ecosystem Open-Weight License and requires a separate commercial license for a model-as-a-service business.
DeepSeek V4.1 Flash's weights use the MIT license, which permits commercial use subject to including the copyright and permission notices in copies or substantial portions. For Kimi, read the Kimi K3.1 Pro license, which provides exceptions for internal use and use through official products or certified inference partners. The MiniMax M3.1 Flash license requires you to display "Built with MiniMax M3.1-Flash".
Before choosing a model
0 of 5 done
Use Requesty's OpenAI-compatible API to run hosted model candidates through the same agent. Usage analytics track cost, latency and cache savings by model, while fallback policies handle provider failures.
Requesty is a hosted gateway. If inference must remain in the EU, select an EU-region model and review the EU routing documentation.
Frequently asked questions
- Which open-weight model is best for agents on a budget?
- DeepSeek V4.1 Flash is a budget candidate for coding agents: it solved 77% of Requesty's 129 backend tasks at $0.41 per solve in the November 2026 comparison.
- Which open-weight model solved the most tasks in Requesty's comparison?
- Kimi K3.1 Pro solved 81% of Requesty's 129 backend coding tasks, the highest success rate among the 11 open-weight models in this comparison.
- Is DeepSeek V4.1 Flash a good model for coding agents?
- DeepSeek V4.1 Flash solved 77% of Requesty's 129 backend coding tasks at $0.41 per solve. It is a candidate for teams weighing task success against model cost.
- Is GLM-5.3 Flash available under a permissive license?
- The official GLM-5.3 Flash model repository publishes an MIT license.
- Can businesses use Mistral Large 4 commercially?
- Mistral Large 4's model card permits commercial and noncommercial use under the Mistral Ecosystem Open-Weight License. A model-as-a-service business requires a separate commercial license.
- Does a fast model turn mean a fast completed task?
- A fast model turn does not determine total task time. In Requesty's benchmark, GLM-5.3 Flash averaged 1.0 second per model turn and 11.4 minutes per task.
- OCT '26
The Requesty coding benchmark: 25 models on 129 real backend tasks
We ran 25 models through the same coding agent on 129 private Go backend tasks with hidden tests. Here is who solves the most, who is fastest and who needs the fewest tokens.
- JUN '26
Best AI Coding Model (2026): Benchmarks, Cost, and Real World Performance
Claude Fable 5, GPT-5.5, Claude Opus 4.8, Gemini 3.5 Flash, DeepSeek V4, and Kimi K2.7 Code all claim top coding performance in 2026. This guide compares them on SWE-bench, Terminal-Bench, FrontierCode, cost per million tokens, and real-world agentic coding tasks so you can pick the right model for your workload.
- JUL '26
Same weights, twelve providers, 4.8x price gap: the provider variance problem
On July 20 the same DeepSeek V4 Pro weights were listed across twelve providers from $0.44 to $2.10 per million tokens. Open weights did not commoditize inference, they moved the variance from the model to the provider. Here is how to route for it.