AI Model Tracker: Efficiency Updates, July 30–Aug 1

Published August 2, 2026 · Research window: July 21 – August 1, 2026 (UTC) · Author context: Essa Mamdani, AI/full-stack developer

This week the frontier did not move up; it moved sideways at lower cost. The dominant signal from July 30 to August 1 is a price-performance squeeze across the workhorse tier: OpenAI cut GPT-5.6 Luna and Terra prices, Google shipped a cheaper, more token-efficient Gemini 3.6 Flash, and DeepSeek turned its V4-Flash preview into a public beta with a big agentic-benchmark jump and OpenAI-compatible Responses API support. None of these releases claim a new reasoning crown. All of them claim that agents can now run more steps per dollar without falling off a quality cliff.

OpenAI GPT-5.6: the price cut is the product update

OpenAI’s July 30 post frames the change as passing efficiency gains to customers. The numbers are straightforward: GPT-5.6 Luna output drops from $6.00 to $1.20 per million tokens, and input drops from $1.00 to $0.20. Terra output drops from $15.00 to $12.00, with input from $2.50 to $2.00. Sol pricing is unchanged at $5.00/$30.00, but it gains a Fast mode that OpenAI says runs up to 2.5× faster at 2× price.

The screenshot proves the pricing change is live in the API docs. The most interesting row is Luna. OpenAI explicitly compares it to “frontier-class models from a year ago” at roughly six cents on the dollar per task and nine times the speed. On Agents’ Last Exam, OpenAI claims Luna outperforms Claude Fable 5 at an estimated cost per task nearly 99% lower.

That vendor-reported cost-per-task comparison mixes model behavior, eval harness, and prompt design into one number. Treat it as a directional signal that Luna is no longer a throwaway routing tier, not as a reproducible benchmark. Use Luna for high-volume, well-specified subtasks; keep Sol for ambiguous planning work.

Gemini 3.6 Flash: fewer tokens, fewer dollars, better scores

Google’s July 21 release is the clearest efficiency story of the week. Gemini 3.6 Flash is priced at $1.50/1M input and $7.50/1M output, down from 3.5 Flash’s $9.00/1M output. Google says the model uses 17% fewer output tokens on the Artificial Analysis Index, and up to 65% fewer on DeepSWE. The benchmark deltas are concrete: DeepSWE rises from 37% to 49%, MLE Bench from 49.7% to 63.9%, and OSWorld-Verified from 78.4% to 83.0%.

This chart is defensible because all six values come from the same Google announcement and compare a model directly to its predecessor. The takeaway is that the cost reduction is compounded by a real token-verbosity drop. Computer use is now a built-in client-side tool in the Gemini API and Gemini Enterprise, which matters for anyone building browser or desktop agents.

Gemini 3.5 Flash-Lite is the companion release: $0.30/1M input, $2.50/1M output, 350 output tokens per second, and stronger agentic scores than older 3.1 Flash-Lite. It is positioned for high-throughput, low-latency slots like search, document processing, and first-pass routing.

DeepSeek V4-Flash-0731: the open-weight-style API counter-pressure

DeepSeek’s July 31 update turns the V4-Flash preview into a public beta. The architecture is unchanged: 284B total, 13B active per token, 1M context. The change is re-post-training and API packaging. DeepSeek claims the new build beats the V4-Pro-Preview on several agent benchmarks: Terminal Bench 2.1 at 82.7, DeepSWE at 54.4, Cybergym at 76.7, and Toolathlon Verified at 70.3.

Pricing is the headline: $0.14 per million input tokens and $0.28 per million output tokens. That is roughly one-fourth of Luna’s new price and two orders of magnitude below Sol. The API now supports the OpenAI Responses API format and is specifically adapted for Codex, which lowers migration cost for teams already on OpenAI’s tooling.

The screenshot confirms the release is live in the API changelog. The caveat is that several listed scores are internal DeepSeek test sets (DSBench-FullStack and DSBench-Hard), and the public benchmark harness for the code-agent tasks is still marked as “to be released soon.” Treat the public scores as interesting, and the internal ones as unverified.

What the price-performance map looks like now

The table mixes vendor-reported prices with vendor-reported benchmark claims. That is fine as long as you do not treat it as an independent leaderboard. Prices are contractual and current as of August 1. Benchmarks are claims that need independent confirmation before they drive architecture decisions.

What I would pick

For high-volume agentic routing and tool loops, DeepSeek V4-Flash-0731 is now the cheapest per-token option with a 1M context and OpenAI-compatible API. The risk is provider reliability, region availability, and the fact that some of its strongest scores are internal benchmarks. Run a controlled eval on your own task distribution first.

For teams already inside Google Cloud or the Gemini API, Gemini 3.6 Flash is the safest efficiency upgrade because the cost reduction is paired with a real token-verbosity drop and built-in computer use.

For teams embedded in the OpenAI stack, re-benchmark Luna on your own tasks. At $1.20/1M output it is cheap enough to replace Terra for many subtasks, and Fast mode on Sol is useful only when latency is the bottleneck and the budget is already approved.

If you are building evals to decide between these tiers, see AI Coding Agents: Skills, Plugins, MCP, ACP, Memory and Workflows, Structured Outputs for Reliable AI APIs, and OpenTelemetry GenAI Observability. Last week’s tracker covered Claude Opus 5, Grok 4.5, and Kimi K3’s independent cyber assessment.

FAQ

Are these prices comparable across providers?

Yes, as listed prices per million tokens. No, once you factor in context caching, batch discounts, concurrency limits, peak-hour surcharges, and regional uplifts. DeepSeek already warns that peak/off-peak pricing will apply soon.

Is Luna really “frontier-class from a year ago”?

That is OpenAI’s framing, not an independent claim. Use it as a screening hypothesis, then run your own evals.

Why trust Gemini 3.6 Flash’s efficiency claims more than cross-vendor benchmarks?

Because the comparison is against its immediate predecessor on the same benchmarks, reported by the same source. Relative deltas within one vendor’s release are usually more reliable than cross-vendor leaderboard comparisons.

Should I switch from OpenAI to DeepSeek for cost?

Only if your integration surface, data residency, latency, and reliability requirements all survive the switch. The price gap is large, but migration and operational risk are not zero.

Methodology and scope

I searched official vendor pages, API docs, and benchmark sites for updates published between July 21 and August 1, 2026. Vendor pricing and release claims are treated as vendor-reported. Benchmark deltas are treated as vendor-reported unless they come from an independent leaderboard. I kept the two classes separate and did not merge them into a single ranking. I also avoided secondhand write-ups when the primary source was available.

Source notes

• OpenAI: Advancing the price-performance frontier with GPT-5.6
• OpenAI API docs: Pricing
• Google: Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
• DeepSeek API docs: Change Log
• DeepSeek API docs: Pricing

Author context

I write about AI systems and full-stack engineering from the point where product behavior meets production reality. My bias is toward explicit contracts, narrow tool surfaces, and enough telemetry to explain failures without exporting private data everywhere.

If you want help turning this kind of tracker into a production-grade eval harness, I’m available through /hire.