Tinker vs Self-hosted RL Cost Calculator
Pick model, workload, steps, batch size — get a side-by-side breakdown.
Lately, I have been Tinkering a lot and learning about post-training: I can easily build RL pipelines, train on big models that I would have never been able to fit on my tiny Stanford GPUs, plus I got some free credits through the research grant program..
I keep getting the same questions when I try to get my labmates to use it or convince advisors to pay for it:
“Why not just rent a GPU and run verl / skyRL yourself? I can spin up an H100 on Lambda for $2/hr and install verl in 20 minutes — is Tinker actually cheaper, or are you just paying for convenience?”
“How much compute do I actually need for this? Like, before I even think about cost — how many GPUs, how much VRAM, how many FLOPs per step? I don't even have a good mental model for what RL training requires vs. regular fine-tuning.”
“If it wasn't for publication, wouldn't you just do this on frontier models? OpenAI has fine-tuning, Gemini has fine-tuning — how much algorithmic control do I actually need, and is it worth giving that up?”
“What else do I even need to pay for? I want to train on math, kernel writing, SWE agents, tool use — what infrastructure am I missing beyond the GPU bill?”
Inference
$/M tokens
Rollout generation — how yappy is your model? Single-turn vs multi-turn prefill multiplier.
Training
$/M tokens
Forward-backward on rollout data. Pay per token, don't care about MFU.
Convergence
Steps to saturate
The hard one — sigmoid scaling, algorithm choice, base model quality.
Environment
$/sandbox-sec
Zero for math, seconds for code, minutes for agentic tasks.
In fact, I have seen these questions before — just dressed differently each time.
2022 — distributed computing
When I worked on Ray for distributed computing: do you rent servers and run Ray yourself, or pay for Anyscale? (Ray was widely used for pretraining (OpenAI, Cohere), but people didn't pay for the managed layer.)
2023 — inference
When everyone was playing with inference: do I host vLLM locally or pay Together AI / Fireworks? This time the managed layer did win — SemiAnalysis calls it tokenomics.
2026 — post-training
Now we are in the age of post-training. The question is the same: where should the abstraction sit, and what does the economics look like this time?
This post explores Tinker-nomics — the economics of RLVR post-training, through the lens of Tinker, the managed training platform from Thinking Machines Lab. What are the technical pieces, the systems challenges, and the real costs? Just like tokenomics for inference, I want to figure out where the value sits for training.
For inference, we have tokenomics — well-understood pricing models where cost scales with tokens consumed. Inspired by SemiAnalysis's tokenomics model for inference, this guide asks: what does the same analysis look like for post-training? GPU-hours, utilization gaps, phase transitions, LoRA multi-tenancy — the cost structure of reinforcement learning is wildly different from what most people assume.
This interactive guide breaks down RLVR (Reinforcement Learning with Verifiable Rewards) cost anatomy through the lens of Tinker.
In this guide
Motivation
The inference cost war, the need for abstraction, and why post-training economics are different.
RL Systems for LLMs
Three phases of RLVR training: rollout, reward, policy update. Where does the time actually go?
FLOP Accounting
The naive calculation: 2N vs 6N FLOPs, MFU gaps, and why fewer FLOPs can cost more.
LoRA & MoE
How LoRA enables multi-tenancy, how MoE flips the cost from compute to memory.
Efficiency Techniques
Two axes of improvement: fewer samples needed, and each sample cheaper.
Convergence
How many tokens to saturate? Sigmoid scaling laws, algorithm tradeoffs, and the 25% rule.
Training Quality
Why precision matters for RL training, off-policyness limits, and framework tradeoffs.
Abstraction Design
Per-token pricing, four primitives, and why the right abstraction layer changes the economics.
Tinker Multiplexing
Inside Tinker's clock cycles, multi-tenant LoRA scheduling, and why per-token pricing works.
Eval Infra & Sandboxing
Reward verification, sandboxed code execution, latency budgets, and reward hacking.
Cost Model
How the calculator works: roofline model, 4N LoRA FLOPs, token breakdown, known gaps.
Sample Workloads
math_rl, code_rl, harbor_rl, osworld_rl — reference workloads, their characteristics, and cost drivers.
Cost Calculator
Plug in your parameters and compare Tinker vs self-hosted GPU costs interactively.
Takeaway
When to use Tinker vs self-hosted — a decision framework based on workload, scale, and team.