I Read the Speed Chart Again

This is my personal record for understanding markets, not a recommendation to buy or sell any asset.

Source cutoff: the morning of August 14, 2026, Korea time

Today’s briefing put faster models, a new agent protocol, data-platform funding, and autonomous payments on one crowded page. The headlines made it sound as if the AI industry had tied its shoes overnight and started sprinting again.

The official dates and availability told a calmer story. The direction was mostly right, but new announcements were mixed with features already released months earlier. I therefore separated what I could verify from figures that still lacked matching primary documentation. Faster AI does not make the calendar run faster too.

1. GPT-5.6 Sol is fast, but I kept two speed charts separate

OpenAI’s API documentation describes GPT-5.6 Sol as a frontier model with a 1.05-million-token context window and up to 128,000 output tokens. Its separate Fast mode, requested with service_tier: "fast", offers up to 2.5× Standard speed and more consistent latency.

Fast mode carries a premium. A sharp traffic ramp can also downgrade requests to Standard processing, so I need to inspect the returned service_tier rather than assume every request used the faster path.

The supplied briefing described a limited “Ultrafast” preview at up to 14× speed and 750 output tokens per second. Those figures and broad availability were not yet reflected in the official API pages I checked. I did not present the public Fast mode and a limited-preview claim as the same product. Seeing a 300-km/h test car does not shorten my commute immediately.

Sources: OpenAI GPT-5.6 Sol documentation, OpenAI API Fast mode

2. Swappable agent runtimes matter, but the work began in June

Vercel introduced HarnessAgent on June 12 as an experimental common interface for harnesses such as Claude Code, Codex, and Pi. AI SDK 7, announced on June 25, organized a broader harness layer including Deep Agents and OpenCode.

The direction is useful: instead of hard-coding one agent into business logic, I can separate business task → AgentRuntime → actual agent. That makes models and execution environments easier to replace.

I could not confirm the exact claim that an ACP adapter first arrived on August 13 in the same form as the briefing. Vercel also warns that harness packages remain experimental and may introduce breaking changes. Before pouring concrete around a new protocol, I would rather install a modular floor.

Sources: Vercel introduces HarnessAgent, Vercel AI SDK 7

3. AI-written code still needs a license receipt

GitHub announced an enterprise open-source license-compliance public preview on June 30. Organizations can define centralized allow-and-deny policies and catch noncompliant dependencies before a pull request is merged.

I could not confirm the briefing’s exact reduction in unknown licenses from 45% to 24%, so I left that number out. The practical lesson remains clear:

• Review new dependencies in pull requests.
• Generate an SBOM covering direct and transitive packages.
• Automate license allow-and-deny policies.
• Review AI-suggested packages under the same rules as human choices.

“One package will solve it” may be true for the code and still create a fresh calendar for the legal team. The install command is one line; the obligations can be a novel.

Source: GitHub open-source license compliance public preview

4. Agent payments have moved from vision to a test bench

Circle announced Agent Stack on May 11, combining Agent Wallets, Agent Marketplace, and Nanopayments. Nanopayments had already launched on testnet in March, enabling experiments with small per-request USDC payments over x402.

The idea that an agent can purchase an API or another service per call has reached a real product experiment. I still kept that separate from the briefing’s August 13 roadmap, 900-service count, and 99.3% settlement share, which I could not match to the same official source.

Once an agent gets a wallet, the first feature should be a limit: per-call spend, daily total, approved sellers, duplicate-payment protection, and receipts. A company that hands a new employee a corporate card and says “use your instincts” rarely enjoys a quiet month.

Sources: Circle Agent Stack, Circle Nanopayments testnet

Claims I did not present as confirmed

I left the following out as definitive August 13 facts because I could not verify matching primary documentation and terms:

• The same-day launch of Gemini 3.7 Flash across GitHub and Vercel
• The detailed maximum 11× DeepSeek V4-Pro-0813 price increase
• A $190 billion valuation and new $5 billion Databricks round
• GitHub’s 45% to 24% unknown-license figures
• The exact same-day scope of Vercel’s ACP adapter and Grok Build support

These may be important. They simply did not have enough date-and-scope evidence for me to publish them as settled facts. Four sourced dishes are easier on the stomach than a ten-plate news buffet with mystery ingredients.

My one-line takeaway

The important story was not “the fastest model.” It was the need to measure task completion × latency × cost × recoverability together. Models are faster, runtimes are becoming replaceable, and agents are beginning to spend money. One benchmark can no longer describe the whole system.

I put the practical dashboard in AI Technology: Give the Agent a Dashboard.