4 AI drops worth watching: July 22
Google: Gemini 3.6 Flash and 3.5 Flash-Lite trade tokens for price
On July 21, Google released Gemini 3.6 Flash and 3.5 Flash-Lite (alongside a security-specific variant covered below). The headline is efficiency: 3.6 Flash uses 17% fewer output tokens than 3.5 Flash while posting real capability gains, moving DeepSWE from 37% to 49%, MLE Bench from 49.7% to 63.9%, and OSWorld-Verified from 78.4% to 83.0%. Pricing lands at $1.50 per million input tokens and $7.50 per million output for 3.6 Flash, and $0.30 input and $2.50 output for Flash-Lite, which Artificial Analysis clocks at 350 output tokens per second. Both are available through the Gemini API, Android Studio, Google AI Studio, and the Gemini Enterprise Agent Platform.
Fewer tokens per answer at a lower per-token price is a compounding win: it cuts cost twice for any workload that runs Flash at volume.
The take. Flash is the tier that actually carries production traffic, so token efficiency matters more here than a benchmark point on the flagship. The 17% output-token reduction is the number to watch, because it lowers the bill on every request without a price change, and the price did change too. Flash-Lite at $0.30 input is positioned to become the default for high-volume classification and routing, where the reasoning ceiling matters less than throughput and cost. Watch whether the DeepSWE jump holds up in real agent loops; a 12-point benchmark gain on coding tasks is meaningful only if it survives multi-step execution outside the eval harness.
Definitions:
- DeepSWE: A benchmark measuring how well a model completes long, multi-step software engineering tasks.
- OSWorld-Verified: A benchmark measuring whether a model can operate a real computer desktop to finish tasks.
- Output tokens: The text a model generates in its answer, billed separately from and usually more expensively than input tokens.
OpenAI: Presence turns agent reliability into a product
On July 22, OpenAI introduced Presence, an enterprise platform for running agents in production. The framing is explicit: “The challenge for enterprises is no longer proving that AI agents can work, it’s making them reliable enough to do high-value work in production.” Presence wraps an agent in policies that define what actions it can take, guardrails that intervene when it moves outside company boundaries, escalation rules that hand off to humans, and a Codex-powered loop that suggests improvements from production performance. It is available today for voice and chat agents, but through a limited general-availability program: deployments are managed by OpenAI Forward Deployed Engineers and select integrators, and it “is not yet available as a self-serve product.”
The product is not a new model. It is the scaffolding around a model that a regulated enterprise needs before it will let an agent touch a real customer.
The take. This is OpenAI conceding that the hard part of enterprise agents is operations, not intelligence, and selling the operations layer directly. The hand-deployed, engineer-managed rollout is the tell: this is a high-touch enterprise motion, not a developer product, and the economics only work on large contracts. For independent builders the immediate value is muted, but the design is a preview of what production agent infrastructure looks like: policy, guardrail, escalation, and a tuning loop as first-class components. Watch whether Presence ever becomes self-serve. Until it does, the reliability tooling is gated behind OpenAI’s sales team, and the open-source stack has room to build the same primitives.
Definitions:
- Guardrails: Automated checks that block or correct an agent’s action when it violates a defined boundary.
- Escalation: Handing a task off from an agent to a human when confidence or policy requires it.
- Forward Deployed Engineer: A vendor engineer embedded with a customer to build and tune a deployment on-site.
Google DeepMind: Flash Cyber out-finds Opus 4.6 on real vulnerabilities
On July 21, DeepMind introduced Gemini 3.5 Flash Cyber, a lightweight model built on 3.5 Flash and fine-tuned to find, validate, and patch software vulnerabilities. The results are concrete: on the V8 JavaScript engine it found 55 unique confirmed issues, compared to 47 for mainline 3.5 Flash and 36 for Opus 4.6, and it surfaced remote code execution vulnerabilities in about two hours. DeepMind reports competitive performance against much larger models on the CyberGym benchmark and a significant uplift over mainline Flash on Chrome’s production security pipeline. Access is restricted: the model is available exclusively to governments and trusted partners through CodeMender in a limited pilot, expanding over time.
Building the vulnerability-finding capability into a cheap, fast model is what makes it usable at scale: it can run on frequent scans and commit pipelines rather than one-off audits.
The take. The load-bearing claim is that a small, fine-tuned model beats a frontier model at a specialized security task, and DeepMind put numbers on it: 55 versus 36 confirmed V8 issues against Opus 4.6. That is the case for specialization over raw scale, and if it generalizes it reshapes how security teams buy AI. The gated access is the honest part of the announcement; a model that finds real RCE bugs in hours is dual-use by definition, and government-and-partner-only distribution is the containment choice. Watch whether the pilot expands to security vendors, because the moment vulnerability-finding at this quality reaches the open market, both defenders and attackers get the same tool.
Definitions:
- RCE (Remote Code Execution): A vulnerability class that lets an attacker run arbitrary code on a target machine, among the most severe.
- V8: The open-source JavaScript engine that powers Chrome and Node.js, a high-value target for security research.
- CodeMender: Google’s system for using AI to automatically find and fix code security issues.
Thinking Machines: Inkling puts a near-1T model in open weights
On July 15, Thinking Machines released Inkling, an open-weights decoder-only multimodal Mixture-of-Experts model. It runs roughly 1 trillion parameters (975B total, 41B active) across 256 experts, accepts image, text, and audio inputs, and carries a 1M-token context window after training on 45 trillion tokens. Benchmarks include 97.1% on AIME 2026, 77.6% on SWEBench Verified, and 73.3% on MMMU Pro vision. The release ships with day-zero support in transformers, SGLang, vLLM, and llama.cpp, plus quantized checkpoints: an NVFP4 build fits in 600GB of VRAM, and GGUF versions target local deployment.
The active-parameter count is the practical number here: 41B active means inference cost tracks a mid-sized model even though the weights total near a trillion.
The take. Open weights at this scale, with day-zero support across every major serving stack, is the release that matters most to builders this week. The MoE design is what makes it viable to self-host: you pay for 41B active parameters at inference, not 975B, and the NVFP4 checkpoint fitting in 600GB puts it within reach of a single high-memory node. The benchmark line worth trusting is SWEBench Verified at 77.6%, a coding number that is hard to game. Watch adoption in the fine-tuning ecosystem; an open multimodal model this capable becomes a base for downstream specialists the way earlier open releases did, and day-zero llama.cpp support lowers the barrier to exactly that.
Definitions:
- MoE (Mixture of Experts): An architecture that activates only a subset of parameters per token, so a huge model runs at the cost of a much smaller one.
- Active parameters: The parameters actually used to process each token, which determine inference cost, versus total parameters stored.
- NVFP4: A 4-bit numerical format that shrinks a model’s memory footprint so it fits on less hardware.
For your week ahead: Google split Flash into a cheaper generalist and a government-only security specialist, OpenAI moved the hard part of enterprise agents (reliability) into a product, and Thinking Machines put a near-trillion-parameter model in open weights. The frontier is not just getting bigger, it is fragmenting into tiers built for specific jobs. Build against the tier that fits the workload, not the flagship.
More drops at dropwatch.ai. Want them in your inbox? Subscribe below.