Oh My Pi: DeepSeek + GPT-5.6 Luna Agent Stack

A developer has published a working oh-my-pi (omp) setup that assigns different agent tasks to different tools: DeepSeek V4 Flash handles coding, GPT-5.6 Luna handles images, and Google’s Antigravity CLI supplies web search. The post reached Hacker News on August 2.

The experiment is a useful example of capability routing, but it is not a new model launch, an official integration, or proof that the combination is cheaper per completed task. Buyers should separate the architecture from the author’s personal billing setup.

How the three-part agent stack works

Layer	Role in the published setup	Billing or access path	Main caveat
oh-my-pi	Orchestrates the coding session and extensions	Open-source agent harness	Third-party integrations expand the trust boundary
DeepSeek V4 Flash	Primary text and coding model	Metered DeepSeek API key	No native image input in this workflow
GPT-5.6 Luna	Describes pasted images for the text-only model	The author says omp uses their paid ChatGPT account	Subscription entitlements are not the same as API billing
Antigravity CLI	Runs searches requested by the agent	Separate Google CLI session	Search usage, quotas, and data handling remain a separate cost center

The author built an omp extension that invokes agy, the Antigravity CLI command, when the agent wants current web information. That keeps Google authentication inside Google’s CLI rather than logging Google directly into omp. The linked extension is community code and should be reviewed before installation.

Pricing impact: cheap routing still creates multiple bills

The live table and chart above pull current API rates from our canonical pricing data. They show what an API-based reproduction would start from; they do not reconstruct the author’s subscription bill.

We rechecked DeepSeek’s official cache-hit, cache-miss, and output rates and OpenAI’s current Luna rate on August 3. No model or price mutation is required: the canonical pricing dataset and public pricing API remain byte-identical and match those official sources.

DeepSeek bills Flash by input, cache-hit, and output tokens. OpenAI cut Luna’s API price by 80% on July 30 and also reduced how Luna usage counts against paid Codex and ChatGPT Work subscription quotas. The source uses the latter access pattern for vision, while its DeepSeek calls remain metered.

The source describes Flash as no longer in preview, but DeepSeek’s official documentation still calls the 0731 API release public beta. Production procurement should follow the provider status rather than the community wording.

Antigravity search is a third component, so “no separate search provider” does not mean search has zero economic cost. Google’s official pricing page lists an individual plan at $0 per month with basic weekly rate limits and includes the CLI among product features. Google AI Pro and Ultra users receive higher quotas. Eligible paid-plan users can choose whether purchased AI credits cover overages at Gemini Enterprise Agent Platform consumption pricing. That quota-and-credit model is not a stable per-search price.

DeepSeek now warns of a significant overall API price increase, but it has not published the new rates, affected billing items, or effective date. Our live data remains on the current official rate card until DeepSeek publishes and activates the change; see the price increase analysis for details.

Who benefits—and who takes on more risk

This pattern suits developers who want a low-cost coding model without giving up occasional vision and current-information tools. It also avoids paying a premium model for every routine code-editing token when only a small share of turns need images or search.

The tradeoff is operational complexity. A pasted image can pass through Luna before its description reaches DeepSeek, while a search query and result pass through Antigravity. That creates more processors, credentials, logs, failure modes, and potential prompt-injection surfaces than a single-provider agent.

Teams that need consolidated billing, data-residency guarantees, stable support, or auditable routing may lose more in engineering and governance time than they save in token charges.

What developers should do now

1. Reproduce the route on a disposable repository with non-sensitive images and queries.
2. Log every model and CLI invocation, then compare total cost, latency, and accepted changes—not just token rates.
3. Confirm that each authentication method and subscription entitlement permits the intended third-party workflow.
4. Review and pin the omp extension; do not execute a community gist blindly.
5. Treat search output as untrusted data and prevent it from overriding system or repository instructions.
6. Add per-provider budgets and a fallback for vision or search outages.

Labs status: public models included, combined workflow blocked

DeepSeek V4 Flash and GPT-5.6 Luna are already included in AI Pricing Guru Labs. In the latest accepted 49-task run, each answered 48 tasks correctly with zero API errors. The DeepSeek row remains explicitly labeled pre-0731 because the launch-day refresh was rejected after all 49 requests hit insufficient-credit errors; Luna’s cost was recalculated against the accepted token counts after OpenAI’s price cut.

Those public endpoint results do not validate this multi-tool stack. An end-to-end launch run is blocked because the workflow depends on a local third-party harness, a custom Gist extension, account-specific subscription access, and non-deterministic Antigravity search results. The source also provides no pinned repository, task set, tool-call trace, quota use, or acceptance test. We will not combine separate model scores and present them as evidence for the orchestration layer.

For current rates, use our DeepSeek pricing page, OpenAI pricing page, and token calculator. Our earlier DeepSeek V4 Flash 0731 analysis covers the model’s official benchmark claims and verbosity caveat in detail.

For a hosted DeepSeek route, compare Novita’s current model catalog, but verify that the exact 0731 snapshot is served before testing.

Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored Novita link at no extra cost to you. It does not affect this analysis.

Verdict

The omp experiment shows why model routing is moving from abstract architecture diagrams into everyday developer tools. DeepSeek V4 Flash can carry the high-volume coding loop, while Luna and Antigravity fill two capability gaps only when needed.

Treat it as a promising prototype, not a production reference architecture. The winning version is the one that survives real repository tests, complies with every provider’s account rules, and lowers cost per accepted change after all auxiliary calls are counted.

Sources: the original FlashBlaze setup post, its Hacker News discussion, the official DeepSeek pricing documentation, OpenAI’s API pricing and GPT-5.6 price update, oh-my-pi, Google’s Antigravity CLI page, Antigravity pricing, plan and overage documentation, and CLI credit documentation.