We train you a custom model
that keeps improving as your users use it.
34% cheaper than Tinker. You only pay for results.
Change one URL. We learn from every interaction, build evals from your real traffic, train your custom model, and serve it behind the same URL, with the results to prove it. No ML team needed.
- 34% cheaper than Tinker$16,226 vs $24,762 for the training behind our case study
- Full-parameter, not LoRAevery weight trained; Tinker trains LoRA adapters only
- About 3.3 daysfor that job: 2.81B trained tokens on 48 B300 GPUs
- Proven in productionSOC 2 compliant, serving premier US hedge funds
One base URL. Five steps. You make one decision.
Collect
Change one base URL. We pass every call through to the same model and record each task from start to finish, not as separate calls.
base_url = "https://api.evolution-research.com/v1"
Diagnose
We search those tasks for failure modes and show how often each one happens, with real examples.
Decide
You review the issues and confirm which ones to fix.
Improve
We build evals for each confirmed issue from your own data, then post‑train the model and optimize the harness against them.
Ship
A new endpoint behind the same URL, plus the eval results that show what changed.
A better auditing endpoint for a production AI spreadsheet agent
We ran the loop on the agent's production traffic and post-trained the open-weight model behind it. Every comparison runs both models on the same tasks, in the same environment, scored by the same grader.
An average lift of 8.7 points across all 92 tasks it never trained on
Audits of real banking and private-equity models from real users’ requests, none of them in training: not the task, not its workbook, not the conversation it came from. After our training, 67 updates across five runs, the average score rises from 46.6% to 55.3%.
- Every gradeable task in the held-out pool, not a hand-picked subset
- Checked against training by workbook bytes and source conversation
- 3 runs per task, averaged; one frozen rubric grader for both models
On the skills it trained, it reaches the Opus 5.5 range
On the 20 tasks from the same held-out pool where training landed, the trained model scores 57.5%: above Opus 5.5 at low effort (54.4%) and within about 6 points of Opus 5.5 at medium (62.8%) and high (63.6%).
Notice: these are the 20 tasks with the largest lift out of the 92 in result 01, picked after scoring. They show how far training reaches where it lands, not the typical gain; for the average across all tasks, see 01.
- Still never in training: not the task, its workbook or its conversation
- Opus 5.5 run on the same 20 tasks at three effort levels
General ability goes slightly up, never down
Shortcut’s benchmark for releasing a new agent tests general ability, not just auditing. On it, the audit-trained model scores higher than the base model, so specializing cost nothing elsewhere.
- Shortcut’s own benchmark for releasing a new agent
- 57 tasks across modeling, valuation and reporting
Opus 5.5 at medium effort, on an earlier version of the same 57 tasks.
Evaluation tasks are kept out of training, checked by workbook and source conversation. One frozen rubric grader scores every run, and full traces are kept for every rollout. p-values are one-sided paired tests.
The cheapest way to a proven improvement.
Full-parameter training at GPU cost, cheaper than LoRA on Tinker, with results proven on a production agent used by paying customers. Training platforms sell you GPUs and leave the research to you; Applied Compute embeds engineers under a custom contract. Our researchers set up your pipeline and check the first runs, then the automated loop does the rest.
| Evolution Research | Applied Compute | Tinker | Fireworks | Prime Intellect | GPU clouds | |
|---|---|---|---|---|---|---|
| What you get | ||||||
| Proven on a production agent, live with paying customers | ✓ +8.7 pts on 92 held-out audit tasks | Not published | — | — | — | — |
| An improvement delivered, not just compute | ✓ | ✓ | — | — | — | — |
| Full-parameter training, done for you | ✓ every weight trained | Custom contract | LoRA only | LoRA only when managed full-parameter only with your own training loop | — rent GPUs, build it yourself | You build it |
| Built from your own production traffic | ✓ one base URL | ✓ | You collect it | You collect it | You collect it | You collect it |
| Before/after report on your own tasks | ✓ | Not published | — | — | — | — |
| Mostly automated: experts set up and check the first runs, then the loop runs itself | ✓ | Engineers embedded with your team | — | — | — | — |
| Improved model served behind the same URL | ✓ | ✓ | Tinker-trained adapters only, not for heavy production | ✓ dedicated GPUs | Can’t host a full-parameter model | You run it |
| What it costs | ||||||
| Engineers you need | None | None, theirs are billed in | Your ML team | Your ML team | Your ML team | Your ML and infra team |
| The training behind our case study, GLM-5.3-Flash the same job everywhere: 2.81B trained tokens | $16,226 full-parameter training, all included | Custom enterprise contract | $24,762 LoRA training only, you build data and evals | $56,978 LoRA training only, you build data and evals | ≈ $18,900 GPU rental only, you build everything | $26,900 – $67,500 GPU rental only, you build everything |
| One improvement of another open model | $4,253 – $83,173 every model below | Custom quote | Same, at its rates | Its rates | You build it | You build it |
| Hosting a LoRA model, per 1M tokens | from $0.07 / $0.20 in / out, cheapest host for your model | Custom quote | $0.84 / $2.11 GLM-5.3-Flash | Dedicated GPUs only est. from ≈ $30 an hour | No hosting | You run it |
| Hosting a full-parameter model, per B300 GPU-hour | $4.28 if our GPUs are full, $4.99 – $15.00 elsewhere | Custom quote | Not offered | $15.00 | No hosting bare GPUs from $4.99, as available | $7.10 – $17.80 |
| One model copy on 4 GPUs, per month | ≈ $12,500 | Custom quote | Not offered | ≈ $43,800 | No hosting ≈ $14,600 in bare GPUs | ≈ $20,700 – $52,000 |
The GLM-5.3-Flash row prices the same job at every provider: the training behind our case study, 67 updates and 2.81B trained tokens, with 4.19B uncached and 26.62B cached prompt tokens and 0.51B generated tokens, counted from our training logs (trained and prompt tokens exactly, generated tokens from the serving engines’ throughput logs). Ours is the same job on our pipeline: about 3,790 GPU-hours, $16,226 at $5.77 per 1M trained tokens, measured on our current run with startup included. Tinker and Fireworks at their published per-token training rates (both LoRA), assuming their prompt cache hits as often as ours. Prime Intellect and GPU clouds at their listed B300 hourly rates (GPU clouds: Modal $7.10 to AWS $17.80) for the same 3,790 GPU-hours, assuming they run as efficiently as we do. Hosting per month is 4 GPUs × 730 hours. Fireworks’ managed training is LoRA only; full-parameter training there needs your own training loop. Prime Intellect’s docs cover deploying only LoRA adapters from its own retired training service. Applied Compute publishes no prices. Checked 15 Oct 2026.
Your data is your edge. We guard it that way.
We are SOC 2 compliant, and the same platform already serves premier hedge funds whose models and data are their most closely held intellectual property.
Independently audited security controls across our platform and operations.
Built for clients who treat every workbook, prompt and model as proprietary.
Your traffic, your evals and your improved model are used for your work only.
You pay only for results, at GPU cost.
Full-parameter training, the full-performance option, done for you and cheaper than LoRA on Tinker. Before training we agree the success test with you. If the new model passes it, you pay for the tokens it was trained on, at GPU cost; if it does not, you pay nothing for the training. Serving is billed by the GPUs it uses. Finding failures, building evals and the report are included.
The training behind our case study, 2.81B trained tokens, costs about $16,226 at this rate. Measured on our current run: 48 B300s at $4.28 an hour trained 689M tokens in 929 GPU-hours, startup included.
- Full-parameter training: every weight updated, for full performance
- 34% cheaper than LoRA on Tinker: $16,226 vs $24,762 for the same job
- GLM-5.3-Flash
- Hosted on our own GPUs, $4.28 per GPU-hour
- If our GPUs are full: other providers at $4.99 – $15.00 per GPU-hour, passed through
- Diagnosis, evals and before/after report
- Pay only if the model passes the success test agreed before training
For other open models, or when our GPUs are fully booked. The training behind our case study would cost about $24,762 there for GLM-5.3-Flash.
- Any model on Tinker, LoRA training only
- Hosted per token on Tinker at its sampling rates, or on dedicated GPUs
- Diagnosis, evals and before/after report
- Pay only if the model passes the success test agreed before training
A private copy of your trained model on 4+ B300 GPUs; a LoRA adapter is merged into the weights first. This is how we serve GLM-5.3-Flash in production today. $4.28 on our GPUs as capacity allows; if they are full, we host it with other providers at $4.99 – $15.00 per GPU-hour and pass that rate through.
Tinker for adapters trained there, or CoreWeave where that is cheaper (rank 16 or less), passed through. Every model’s rate is in the table below. Tinker does not yet recommend its per-token serving for heavy production traffic. Our own shared per-token GLM-5.3-Flash hosting, about $0.31 / $1.04, comes once we have measured its speed.
- Same URL, no code changes
Every model: what it costs to train, and the cheapest way to host it
GLM-5.3-Flash trains on our own GPUs. Other open models, and overflow when our GPUs are booked, train on Tinker as LoRA adapters at Tinker’s published rates. Hosting goes wherever is cheapest for that model.
| Model | Trains on | Est. one improvement* | Cheapest per-token hosting, per 1M tokens** | Or dedicated, our GPUs*** |
|---|---|---|---|---|
| GLM-5.3-Flash 320B / 18B | Our GPUs $5.77 per 1M trained tokens, at cost | $16,226 | Dedicated now shared per token ≈ $0.31 in · $0.06 cached · $1.04 out, once proven | 4 GPUs ≈ $12,500 a month |
| GLM-5.3-Flash 320B / 18B | Tinker its rates, passed through | $24,762 | Tinker $0.84 in · $0.168 cached · $2.11 out | 4 GPUs ≈ $12,500 a month |
| GLM-5.3 753B / 40B | Tinker its rates, passed through | $83,173 | Tinker $3.78 in · $0.756 cached · $12.15 out | 8 GPUs ≈ $25,000 a month |
| Kimi K2.6 1T / 32B | Tinker its rates, passed through | $67,093 | Tinker $2.205 in · $0.441 cached · $5.49 out | 8 GPUs ≈ $25,000 a month |
| Qwen 3.5 397B 397B / 17B | Tinker its rates, passed through | $53,273 | Tinker $1.65 in · $0.33 cached · $7.50 out | 4 GPUs ≈ $12,500 a month |
| DeepSeek V4.1 Flash 552B / 16B | Tinker its rates, passed through | $21,550 | Tinker $0.60 in · $0.12 cached · $3.60 out | 4 GPUs ≈ $12,500 a month |
| Qwen 3.8 27B 27B dense | Tinker its rates, passed through | $30,469 | CoreWeave $0.40 in · $0.15 cached · $3.00 out | 1 GPU ≈ $3,100 a month |
| Qwen 3.6 35B 35B / 3B | Tinker its rates, passed through | $10,982 | Tinker $0.45 in · $0.09 cached · $1.335 out | 1 GPU ≈ $3,100 a month |
| Nemotron 3.5 Lightning 30B 30B / 3B | Tinker its rates, passed through | $4,253 | CoreWeave $0.07 in · $0.04 cached · $0.20 out | 1 GPU ≈ $3,100 a month |
| GPT-OSS 120B 117B / 5B | Tinker its rates, passed through | $10,121 | Tinker $0.33 in · $0.066 cached · $0.84 out | 1 GPU ≈ $3,100 a month |
| Qwen 3.5 4B 4B dense | Tinker its rates, passed through | $7,422 | Tinker $0.33 in · $0.066 cached · $1.005 out | 1 GPU ≈ $3,100 a month |
Tinker training rates per 1M tokens
| Model | Prompt | Prompt, cached | Generated | Trained |
|---|---|---|---|---|
| GLM-5.3-Flash | $0.84 | $0.168 | $2.11 | $5.58 |
| GLM-5.3 | $3.78 | $0.756 | $12.15 | $14.58 |
| Kimi K2.6 | $2.205 | $0.441 | $5.49 | $15.4 |
| Qwen 3.5 397B | $1.65 | $0.33 | $7.5 | $12 |
| DeepSeek V4.1 Flash | $0.6 | $0.12 | $3.6 | $4.98 |
| Qwen 3.8 27B | $0.6975 | $0.1395 | $5.595 | $7.46 |
| Qwen 3.6 35B | $0.45 | $0.09 | $1.335 | $2.14 |
| Nemotron 3.5 Lightning 30B | $0.184 | $0.0368 | $0.495 | $0.8 |
| GPT-OSS 120B | $0.33 | $0.066 | $0.84 | $2.33 |
| Qwen 3.5 4B | $0.33 | $0.066 | $1.005 | $1.34 |
*The same job as our case study: 2.81B trained tokens, 4.19B uncached and 26.62B cached prompt tokens and 0.51B generated tokens, counted from our training logs, at Tinker’s published list prices for the long-context version of each model (Nemotron at its current limited-time 50% discount), assuming its prompt cache hits as often as ours. Smaller jobs cost proportionally less. On our GPUs one rate covers everything. **Input, cached input and output prices per 1M tokens, at the host’s published rates, passed through: Tinker sampling rates, or CoreWeave’s base-model rates for LoRA. Tinker does not yet recommend its per-token serving for heavy production traffic. ***Estimated B300 GPUs for one private copy, at $4.28 per GPU-hour × 730 hours. Prices checked 15 Oct 2026.
Try it on one of your agents.
A pilot covers one agent. We build evals from your traffic and show you every eval task to confirm before any training, then report the before and after on the evals we agreed on.
Or write to hello@evolution-research.com