Parsec vs Baseten

Baseten serves open and custom models on dedicated GPUs — per-minute billing, fast cold starts, a 99.99% uptime SLA. It lowers the cost of every token it serves; Parsec lowers the tokens the model reads and writes, and reaches the frontier closed APIs Baseten can't host — where the measured −39% happened. Compare · Dasein · 2026-07-15. Chart of quality against cost per task: five model pairs, each with an arrow showing Parsec cutting its cost at constant quality — Sonnet 4.6 from $2.58 to a measured $1.45, Grok 4.5 from $2.49 to an estimated $1.52, Opus 4.8 from an estimated $4.30 to $2.40, GPT-5.5 from $5.07 to an estimated $3.09, and Fable 5 from $11.80 to an estimated $7.20. Key takeaways: Baseten bills dedicated GPUs by the active minute. H100 80GB at $0.10833/min ($6.50/hr), with scale-to-zero, fast cold starts, and a 99.99% uptime SLA; open-model APIs start at $0.10/1M input (GPT-OSS-120B). $1.45 per solved task, against $2.58 uncompressed. 39% off the bill on 100 graded SWE-bench Verified tasks, while solving more (62/100 vs 57/100). The two layers multiply. Baseten lowers the price per token served; Parsec lowers how many tokens the model reads and writes — on a model hosted on Baseten the savings compound. Serving optimization can't touch the frontier half. The measured −39% landed on a closed Anthropic API Baseten does not host; on a split stack, Parsec covers both branches, and every frontier pair drops ~40% at constant quality. What Baseten is: An inference platform for open-source, custom, and fine-tuned models. Dedicated deployments run on single-tenant GPUs billed by the active minute — H100 80GB at $0.10833/min — with scale-to-zero idle replicas, cold starts it advertises in "seconds" for most models, a 99.99% uptime SLA, and SOC 2 Type II / HIPAA compliance. Its Model APIs serve open weights per token: GPT-OSS-120B at $0.10/$0.50, GLM 4.7 at $0.60/$2.20, DeepSeek V4 at $1.74/$3.48 per million. One trade-off from an independent comparison: its managed H100 lists at $6.50/hr against $1.79 on a commodity GPU cloud — the premium buys the autoscaler, cold starts, and SLA. What Parsec is: Dasein's context-compression layer for coding agents. It sits in the request path and curates the working context that grows every turn — tool results, file reads, transcript — so the model reads what it needs and no more, without churning the provider's prompt cache. It is provider-agnostic: frontier APIs and self-hosted open models alike, nothing else in the stack changes. Request-path diagram: a coding agent feeds Parsec, which curates the working context every turn, then forks to two branches — Baseten serving open and custom models per GPU-minute, and frontier closed APIs (Anthropic, OpenAI, xAI) where the measured −39% happened. Parsec cuts the tokens the model reads and writes on both branches; Baseten optimizes cost per token served on the open-model branch only. The measured result: $1.45 per solved task · 62/100 solved vs 57 baseline · −39% total cost · −25% wall-clock. One fixed agent (headless Claude Code on claude-sonnet-4-6), 100 SWE-bench Verified tasks, official Docker grader, cache-aware pricing. Cost per task, side by side: Claude Sonnet 4.6 + Parsec — 79.6%† SWE-bench Verified — $1.45 — measured, Code-Compression Bench. Grok 4.5 + Parsec — 86.6% — estimated $1.52 — −39% off the measured $2.49. Claude Opus 4.8 + Parsec — 88.6% — estimated $2.40 — bench dollars × Opus/Sonnet list price. Grok 4.5 (Grok Build) — 86.6% — $2.49 — measured, Artificial Analysis. Claude Sonnet 4.6 — 79.6%† — $2.58 — measured, Code-Compression Bench. GPT-5.5 + Parsec — 82.6% — estimated $3.09 — −39% off the measured $5.07. Claude Opus 4.8 — 88.6% — estimated $4.30 — bench dollars × Opus/Sonnet list price. GPT-5.5 (Codex) — 82.6% — $5.07 — measured, Artificial Analysis. Claude Fable 5 + Parsec — 95.0% — estimated $7.20 — −39% off the measured $11.80. Claude Fable 5 (Claude Code) — 95.0% — $11.80 — measured, Artificial Analysis. † Anthropic-published; every other score is Vals AI SWE-bench Verified on one shared harness. These are the frontier closed models a team pairs with self-hosted open ones — the half of a split stack Baseten's serving layer does not reach. Compression, not the model choice, drives the cut: each model's cost falls about 40% with quality unchanged, and Sonnet with Parsec also solves more tasks than the uncompressed baseline. When to choose what. Add Parsec when: Tasks are graded by outcome — CI has to pass, the patch has to merge, and a failed task costs more than its tokens. The agent bill is scaling with adoption and the model is staying put — a measured −39%, $1.45 per solved task on Sonnet 4.6. Part of the stack is a frontier closed API (Claude, GPT, Grok) Baseten does not host — exactly where the measured −39% landed. Open models hosted on Baseten still carry a growing context — Parsec curates the token count Baseten prices. Baseten already covers you when: The workload is a self-hosted or fine-tuned open model that needs production serving — dedicated GPUs, autoscaling, scale-to-zero. Cost per token served is the lever — per-minute GPU billing and 99.99% uptime are the product; open-model APIs start at $0.10/1M input. Cold-start latency and burst traffic are the constraint — the autoscaler and fast cold starts are built for exactly that. Compliance and isolation are hard requirements — SOC 2 Type II, HIPAA, and single-tenant or self-hosted deployments are in the box. FAQ. Does Parsec replace Baseten, or run in front of it? It runs in front of it. Parsec's mechanism — fewer, better tokens per call — is model- and host-agnostic, so it sits in the same request path in front of whatever serves the model, Baseten included. On a frontier closed API Baseten doesn't host, Parsec is where the savings come from at all. Baseten already cuts our serving cost — what would a compression layer add? Price per token served and tokens per task are different levers. Baseten lowers what each token costs; Parsec lowers how many tokens the model reads and writes, so on a model you host on Baseten the two compound. Baseten serves only open and custom models, and the measured −39% landed on a closed Anthropic API it does not host — on that half of a split stack, Parsec is the only lever. Where do these numbers come from? Baseten's per-minute GPU rates, per-token Model API prices, and 99.99% uptime SLA are its own published pricing, retrieved 2026-07-15. The frontier per-task costs — Grok 4.5 $2.49, GPT-5.5 $5.07, Fable 5 $11.80 — are Artificial Analysis's independent measurements. Parsec's $1.45 and the baseline's $2.58 come from the Code-Compression Bench run described above. Run the numbers yourself: The benchmark, methodology, and per-task data are open at https://github.com/daseinlabs/code-compression-bench Wiring Parsec into an existing request path takes one change, and the number that matters is the one measured on your own traffic at https://daseinlabs.ai Sources: Baseten, cloud pricing (dedicated GPU per-minute rates, Model API per-token prices), retrieved 2026-07-15 — https://www.baseten.co/pricing/ Baseten, inference platform homepage (99.99% uptime, fast cold starts, open/custom/fine-tuned model support), retrieved 2026-07-15 — https://www.baseten.co/ Baseten vs DeepInfra H100 price comparison ($6.50/hr vs $1.79/hr), 2026 — https://www.morphllm.com/comparisons/baseten-vs-deepinfra Artificial Analysis, Coding Agent Index (native harness per model), 2026-07-08 — https://artificialanalysis.ai/ Vals AI, SWE-bench Verified leaderboard (same harness across models), 2026-07-09 — https://www.vals.ai/benchmarks/swebench Code-Compression Bench, run 2026-07-04 — sponsored and operated by Dasein, identical methodology for every arm including Parsec's — https://github.com/daseinlabs/code-compression-bench