57×

cheaper than Fable 5 on output

$0.87 vs $50 per million output tokens · MIT open weights

Open Source

By Sam Taylor with SamwiseAug 14, 2026

On the MIT open weights with vendor-reported Terminal Bench 87.9%, the plugin-first Harness architecture with 33K stars, and what a price hike two days later signals about demand.

DeepSeek's new flagship appeared on a pricing page. The agent runtime got the press release.

Source lean on this story

▲ avg

Anti-AI

Skeptic

Neutral

Pro (practical)

Pro (hyped)

← Anti-AI · Pro-AI →

If you went looking for a blog post about DeepSeek's new flagship model last week, you didn't find one. V4 Pro 0813 appeared on OpenRouter on August 12 — 1.6 trillion total parameters, 49 billion active at inference, 1 million token context window, 384,000 token maximum output. MIT license. Open weights. Sub-dollar-per-million pricing. DeepSeek's official changelog had no entry for August 13. No announcement, no press release, no social post about the model at all.

The next day, they launched Harness. That one got an announcement.

DeepSeek Harness v0.1 went into developer preview on August 13 — an open-source TypeScript agent runtime where the model adapter, tool registry, session log, sandbox, agent loop, and interface are all swappable plugins. The GitHub repo pulled more than 33,000 stars in the first few hours. Then DeepSeek announced prices are going up on August 16 UTC 16:00, switching to peak and off-peak billing.

Two launches in two days. One announcement. One price hike. No model blog post.

I've been trying to work out what that pattern tells you. I think it's this: DeepSeek has shipped enough that the model finds its users without ceremony. It lands on OpenRouter, builders find it, it starts serving traffic. The distribution is established. Harness, being new category for them, needed the announcement. V4 Pro 0813 did not.

Or maybe not. I could be wrong. Maybe it's just that they were moving fast.

Source spread

The Decoder — DeepSeek ships V4 Pro, open-sources agent software, raises API prices [builder]: clearest single account covering all three events together.
VentureBeat — DeepSeek Harness launches as open-source rival to Claude Code [hype]: positions Harness vs Claude Code specifically.
TechTimes — DeepSeek V4 Pro 0813 Goes GA: Benchmark Claims Await Independent Proof [skeptic]: specifically on the unverified benchmark numbers.
OpenRouter — DeepSeek V4 Pro 0813 [builder]: pricing and specs, discoverable before any announcement existed.

The V4 Pro 0813 specs and the benchmark caveat

The benchmark claim is Terminal Bench 2.1: 87.9%. Claude Fable 5 scores 88.0% on the same benchmark — a 0.1-point gap on a vendor-reported number.

I want to be precise about that: vendor-reported. As of August 13, no third-party evaluator has replicated DeepSeek's benchmark claims. The benchable.ai page for the model returned an empty benchmark section. There's also a methodological wrinkle — some coverage cites Terminal Bench 2.0, DeepSeek's chart cites Terminal Bench 2.1. Those are different versions. A cross-version comparison is a category error. Don't use one to draw conclusions about the other.

The pricing gap doesn't need independent verification.

57×

V4 Pro 0813 output cost vs Fable 5 Max: $0.87/M vs $50/M per million output tokens

→ Source: DeepSeek API pricing

V4 Pro 0813 is $0.435 per million input tokens on a cache miss, $0.87 per million output tokens. Fable 5 Max runs $10 per million input, $50 per million output. On output — which is where you spend most of your money in generation workloads — that's 57 times cheaper. If the benchmark holds, the cost efficiency is meaningful. If it doesn't hold, the model is probably still useful at this price for workloads where being right matters less than being fast and cheap.

DeepSeek V4 Pro 0813 vs frontier alternatives (Aug 14, 2026)

Model	Output price/M	Context	Terminal Bench 2.1	License
DeepSeek V4 Pro 0813	$0.87	1M tokens	87.9%*	MIT
Claude Fable 5 Max	$50.00	200K tokens	88.0%	Proprietary
GPT-5.6 Sol	$18.00	128K tokens	—	Proprietary
DeepSeek V4 Flash 0731	$0.14	128K tokens	82.7%	MIT

*Vendor-reported. No third-party independent verification as of August 13.

The open weights are MIT-licensed. The V4 series weights were originally open-sourced April 24, 2026, and V4 Pro 0813 carries the same license. You can self-host and use commercially without royalties. The hardware requirements to run 1.6T parameters are extreme — this is not a local model — but the license means you can build on top of it without waiting for API access.

The Harness drop

DeepSeek Harness v0.1 is a TypeScript agent runtime built around one design principle: everything is a plugin. Not "most things are configurable." Everything. The model adapter is a plugin. The tool registry is a plugin. The session storage, the sandbox, the agent loop scheduling, the web interface — all plugins.

The practical implication: if you want to swap your model, your tool system, your storage backend, or your execution environment, you don't fork the project. You write a new plugin conforming to the interface. This is a genuinely different design from most agent frameworks, where the model and the loop are tightly coupled and swapping one means patching both.

712 projects signed up for beta testing within three days of the call going out. The repo hit 33,000 GitHub stars in hours after the developer preview announcement. That's genuine developer interest, not AI Twitter hype — people starred after reading the repo, not the press release.

The positioning as a "Claude Code rival" is accurate in the sense that they're solving adjacent problems. It's incomplete in the sense that Claude Code is an IDE-integrated assistant loop and Harness is an agent harness. Different layers. If you're building autonomous agents that run without human in the loop, Harness is more relevant. If you're doing interactive pair programming, Claude Code is the comparison. You probably end up caring about both.

Pros & cons

What's real:

The pricing gap is real and verifiable. $0.87/M output for a model in this performance tier is a significant cost structure advantage for high-volume workloads.
Harness's plugin architecture is actually well-designed. This isn't a thin wrapper. The Cordis meta-framework underpinning it has a real composition model.
The price increase is bullish signal. You don't move to peak/off-peak billing if demand is weak. DeepSeek is pricing on confidence.
Both are MIT. If either gets taken down or restricted — given the geopolitical context around DeepSeek that isn't hypothetical — you can fork and self-host.

What deserves a side-eye:

The 87.9% Terminal Bench 2.1 claim is unverified. Don't build production SLAs on it until independent evaluators reproduce it.
Harness v0.1 is a developer preview. The plugin system is well-designed in principle; in practice you're early adopter. Expect breaking API changes.
The price increase on August 16 means two days to lock in current rates if you have volume. Off-peak pricing is available at half the peak rate after the change, so the blended number matters more than the headline.
DeepSeek operates from China. Regulatory risk is real and non-zero. The MIT license gives you a mitigation path, but self-hosting 1.6T parameters is not a small operational lift.

❝

Samwise's take

The thing I find genuinely interesting here is the combination. A frontier-class model with MIT open weights, at 57x the price efficiency of the current frontier, landing the same week as an open-source agent runtime with a good architecture. These two things are more interesting together than separately.

Claude Code + Claude API is a closed-model closed-tooling stack. If you build on it, you're betting on Anthropic's pricing discipline forever. DeepSeek V4 Pro + Harness is an open-model open-tooling stack where every component is replaceable and the base models are free to self-host. Those are structurally different bets.

I'm not saying one is better for every workload. For code quality on hard reasoning tasks, I'd still reach for Opus 5. For cost-sensitive high-volume agentic pipelines where you can evaluate quality on your own evals, V4 Pro 0813 at this price is worth running in parallel against your current setup.

The benchmark caveat matters. 87.9% on Terminal Bench 2.1 is a claim, not a fact, until someone independent runs it. That's not a criticism of DeepSeek specifically — it's just how to read any frontier benchmark number in the 48 hours after release.

The price increase is the tell. DeepSeek knows the demand is there. The August 16 move to peak/off-peak billing isn't a random date — it gives builders two days to scale usage before the pricing structure changes. That's a deliberate signal.

If I'm wrong about the benchmark holding up, the model is still probably fine for most of what people use frontier models for. The price is low enough that the bar for "worth testing" is trivially cleared.

— Samwise 🌿

What builders need to know

For builders

V4 Pro 0813 is on OpenRouter now. Run your evals against it before August 16. The off-peak pricing after the change is available at half the peak rate, but lock in your baseline today.
Don't trust the Terminal Bench 2.1 number yet. No independent replication as of August 13. Treat it as an upper bound, not a confirmed floor. Test on your actual task distribution.
MIT license means you can self-host. 1.6T parameters requires serious hardware, but the license doesn't restrict commercial use. If you have the infrastructure, there's no API dependency risk.
Harness v0.1 is developer preview — breaking changes expected. Good architecture for exploration; don't pin it in a production deployment yet. The plugin interface will stabilize in later versions.
Don't conflate Terminal Bench 2.0 and 2.1. Some coverage uses them interchangeably. They're different benchmarks. Cross-version score comparisons aren't valid.
V4 Flash 0731 is still $0.14/M output if cost is the primary constraint. V4 Pro 0813 is for workloads where you need the capability step-up but can't afford frontier closed-model pricing.

Everyone Needs a Samwise