
The best model for writing code is not the strongest one — it is the one whose price and speed fit your workload. Going by effective price and the last 24 hours of traffic on this platform (data as of November 17, 2026): make gpt-5.6-terra or deepseek-v4.1-flash your daily driver. Both logged over three thousand real calls with a 100% success rate, and a month of a thousand coding requests costs $2.85 and $1.66 respectively. When raw output speed matters, reach for gemini-3.8-flash (measured at 111.4 tokens per second, the fastest here). Keep claude-opus-5-5 for the hard refactors. The worst value is kimi-k2.7-code and glm-5.3, both stuck at the 50% price ceiling — the same work costs seven to nine times more.
Three things have to hold at once: it gets the code right (success rate), it gets there fast (output rate and total latency), and it does not cost much (effective price times your volume). Rank-only benchmarks miss the point, and price-only shopping lands you on something too slow to use.
One detail worth flagging: most prices here are the official rate multiplied by a seller's discount, and sellers set their own. The same model can differ several times over between sellers. Every number below is the effective price — what actually leaves your account, not the list price.
These figures come from real requests through this platform's gateway over the past 24 hours, not from vendor benchmark pages. Only one model sits below 99% success: gpt-6-astra at 98.19%, flagged as degraded.
| Model | Calls (24h) | Success | P50 latency | P50 output rate |
|---|---|---|---|---|
| gpt-5.6-terra | 3,902 | 100% | 6.5s | 35.1 tok/s |
| deepseek-v4.1-flash | 3,247 | 100% | 8.3s | 32.4 tok/s |
| gemini-3.8-flash | 2,371 | 100% | 6.3s | 111.4 tok/s |
| gpt-6-astra | 276 | 98.19% | 21.4s | 12.1 tok/s |
| gpt-6.1-sol | 39 | 100% | 80.3s | 37.0 tok/s |
| claude-opus-5-5 | 32 | 100% | 11.6s | 77.7 tok/s |
| claude-sonnet-5-5 | 19 | 100% | 14.9s | 72.9 tok/s |
| glm-5.3 | 4 | 100% | 2.7s | 47.3 tok/s |
Two readings matter. First, the busiest model is not necessarily the smartest, but it is the steadiest — gpt-5.6-terra and deepseek-v4.1-flash together handled more than seven thousand calls with zero failures, which is what production cares about most. Second, do not trust small samples. claude-opus-5-5 has 32 calls and glm-5.3 has four; those rows prove the model works, not that it outruns anything.
One anomaly deserves a note: gpt-6.1-sol shows an 80.3-second P50, twelve times slower than its sibling gpt-5.6-terra, even though its output rate is fine at 37 tok/s. That pattern usually means long reasoning chains or queueing on long tasks. Trial it with a small slice of traffic before trusting it.
For the same model, the official rate and the effective price here often differ a hundredfold. Sorted by effective input price (USD per million tokens):
| Model | Official (in/out) | Discount | Effective (in/out) |
|---|---|---|---|
| gpt-6.1-terra | $1 / $5 | 97% off | $0.0297 / $0.1485 |
| gpt-6.1-sol | $2 / $10 | 98% off | $0.0396 / $0.198 |
| deepseek-v4.1-flash | $0.2987 / $1.1949 | 86.8% off | $0.0394 / $0.1577 |
| gpt-6-sol | $2 / $10 | 98.3% off | $0.033 / $0.165 |
| gemini-3.8-flash | $0.75 / $3.75 | 92.3% off | $0.0578 / $0.2888 |
| gpt-5.6-terra | $2 / $12 | 97% off | $0.0594 / $0.3564 |
| claude-sonnet-5-5 | $2 / $10 | 95% off | $0.099 / $0.495 |
| claude-opus-5-5 | $4 / $20 | 95% off | $0.198 / $0.99 |
| kimi-k2.7-code | $0.9709 / $4.0329 | 50% off | $0.4854 / $2.0164 |
| glm-5.3 | $1.195 / $4.183 | 50% off | $0.5975 / $2.0915 |
Watch the unit: a 0.3 discount means you pay 3% of list, while 5.0 means you pay half. That is how gpt-6.1-terra looks three times pricier than deepseek-v4.1-flash on the official sheet yet costs less in practice.
In the other direction, kimi-k2.7-code and glm-5.3 sit at the 5.0 ceiling — the platform's hard limit, sellers may not quote above 50% of list. A model parked at the ceiling usually means thin supply. Their effective input price runs 12 to 15 times deepseek-v4.1-flash.
Assume a moderate workload: 33 requests a day, each with 30,000 input tokens (code, stack traces, requirements) and 3,000 output tokens. Over a month that is 30 million input plus 3 million output tokens. This is an estimate, not measured traffic, but the ratio mirrors real coding calls:
From $1.66 to $24.20 — a 14.6x spread for the same coding work. The gap is not about capability. It is about discount: deepseek-v4.1-flash is quoted at 86.8% off, glm-5.3 at only 50%.
Daily driver (functions, bug fixes, tests): gpt-5.6-terra or deepseek-v4.1-flash. The first handled 3,902 calls with zero failures at a 6.5-second P50; the second handled 3,247 with zero failures at the lowest monthly cost, $1.66. Run both for a week and let the real invoice decide — that beats any benchmark write-up.
Speed tier (interactive edits, bulk refactors): gemini-3.8-flash. At 111.4 output tokens per second it is the only model in this group above a hundred, with a 6.3-second P50. Note its introductory price only holds through December 31, 2026; after that input goes from $0.75 to $1.50 and output from $3.75 to $7.50, so your effective price doubles with it.
Fallback tier (large refactors, cross-file changes, genuinely hard bugs): claude-opus-5-5. Thirty-two measured calls with zero failures at 77.7 tokens per second, a 1M-token context and 128K max output — enough to swallow a whole module. At $8.91 a month, reserved for hard problems, it will not wreck the invoice.
I would not make kimi-k2.7-code or glm-5.3 the default. Both sit at the 50% ceiling, both cost over $20 a month, and both have single-digit measured samples. Use them for what they are good at: Kimi handles Chinese comments and technical prose well, and GLM's 2.7-second P50 suits fast inline completion.
Treating list price as your cost. A $10-per-million model at 99% off costs $0.01; a $1 model at 50% off costs $0.50. The first is fifty times cheaper. Budget with effective prices.
Ignoring output pricing. In coding, output volume is roughly a tenth of input, but output is priced five to six times higher, so it typically carries a third of the bill. Working through the table above model by model, output accounts for 26% to 38% of the cost. Input price alone tells you very little.
Reading small samples as findings. A model with single-digit calls in 24 hours may be newly listed, or may have a shaky supply. In production, prefer models with thousands of calls and a 100% success rate instead of using your own workload as the test bench.
Open the model marketplace, filter to coding models, and check how many sellers each one has and what discount they quote — discounts move daily, so do not hardcode the numbers above. Then create two keys, one per price tier, and run the same task through both for a week. Finally confirm the live rate on the pricing page before you commit.
If you remember one line: choosing a coding model is price divided by speed, multiplied by success rate. It is arithmetic, not a leaderboard. You got it right when the monthly bill is tens of dollars and the code still runs.
GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer
Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi
Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload
Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa
Hailuo 2.3-fast has the lowest effective price (about $0.00185/s), Kling 3.0 offers the fullest range up to 4K, and Seedance 2.5 costs most but render