OneHuman Watchdog: August 2026 — The Loudest Number Was Never the Real One

In August 2026, six OneHuman reports covered five of the eight AI tools this platform tracks. Four of them share a shape: a vendor disclosed a number, and the number that got repeated in headlines — including, once, our own — was not the number that actually determined what a user or developer paid. The other two share a different shape: the same industry, in the same month, treated a price commitment and a usage-limit commitment as different kinds of promise, keeping one and only ever hoping about the other.

Neither pattern is a story about any single company lying. Every case below is a case of a company saying something true, in public, that a reader had to work to translate into the number that mattered.

---

Pattern 1: The Loudest Number Isn't the Real One

DeepSeek's API price increase, effective August 16, is being reported everywhere as a 1,114% jump. That figure is accurate — and it applies to cache-hit input pricing, which starts from a fraction of a cent and still costs under 5 cents per million tokens at peak after the increase. The number that actually moves a real invoice is output pricing, up 355% for V4-Pro and 371% for V4-Flash. A coding session that ran about $2 in output charges before August 16 runs roughly $9 today at peak hours, $4.50 off-peak. DeepSeek had pre-warned this was coming on August 6, with no figure and no date attached — a disclosure that satisfied the letter of "we told you" while giving developers nothing to plan against for ten days.

OpenAI's July 30 cut to GPT-5.6 Luna's API pricing ran the same shape in reverse. An 80% price cut sounds like a win, and coverage largely treated it as one. The cut was measured against a price OpenAI itself had set three weeks earlier, at launch. Measured against the market, new Luna's rate still loses to DeepSeek V4-Flash on both input (43% higher) and output (over 4x higher) — a fact the 80% headline doesn't surface, because it wasn't a comparison against the field, it was a comparison against OpenAI's own recent past.

Perplexity's case is smaller in dollar terms but the same shape: its free-tier "Research" allowance (renamed from "Deep Research" in 2025) dropped from 5 queries a day to 1 a month — a cut of more than 99% in practical terms. OneHuman's own July 19 report on the same feature, published four days after the cut had already taken effect, had the 5-per-day figure and missed that a smaller number now sat behind the same name. We corrected this in the open on August 5, the same day we found it, rather than leaving the stale figure standing.

Tools affected: DeepSeek, ChatGPT, Perplexity. Consumer verdict: Before repeating any AI vendor's percentage, ask what it's being measured against — the market's actual cheapest option, or the vendor's own most recent, possibly inflated, price.

---

Pattern 2: A Promise on Price, a Hope on Everything Else

Anthropic did two things in August that look similar and aren't. Around August 11, it canceled a scheduled September 1 price increase on Claude Sonnet 5 — the $2/$10 per-million-token introductory rate becomes the permanent standard rate, stated plainly in its own pricing documentation, no hedge attached. Eight days later, on August 18-19, it extended Claude Code's "temporary" 50% weekly usage-limit boost for the fourth time since the boost launched May 13. The extension came with a stated hope: "We hope to make this a permanent change to our plans, but strong demand for our models means that capacity may be tight over the coming weeks." Four renewals, three and a half months, and the hope has never once become a date past the current deadline of August 31.

OpenAI's case is starker: no hope offered at all. Since GPT-5.6 launched July 9, users on paid tiers have documented — first in a July 23 developer-forum thread, then again August 12, then in an August 21 case filed by a ChatGPT Business admin with request IDs and HAR files, then again in a thread posted the morning of August 22 — that requests for GPT-5.6 are sometimes silently answered by the cheaper GPT-5.5-mini instead, confirmed via a resolved_model_slug field in the network response. As of this report, OpenAI has not acknowledged the routing behavior anywhere it publishes. A companion industry scorecard, published by data-quality firm Pebblous on August 14 and explicitly self-described as preliminary, found none of nine major AI vendors could let an outside party verify that the model documented is the model actually served — OpenAI scored highest of the nine, at 23 of 37 possible points, which is itself the finding: the best performer in the industry still fails most of the checks.

Tools affected: Anthropic, ChatGPT. Consumer verdict: A vendor's willingness to put a firm, dated commitment on one thing and not another is itself information. Anthropic's own contrast this month shows both are possible from the same company in the same month — the choice of which gets a date is not accidental.

---

Tool-by-Tool: What Changed This Month

Tool	Key Move	Impact	User Action
ChatGPT	Luna API cut 80% (still not market-cheapest); GPT-5.6 model-routing bug unacknowledged since Jul 9	High	Compare Luna against DeepSeek V4-Flash directly, don't assume the cut makes it the cheap option; check your own network tab if a response feels off
DeepSeek	API moved to peak/off-peak pricing Aug 16 — output up 355-371%, cache-hit headline (+1,114%) trivial in absolute terms	High	Budget around peak-hour output pricing, not the headline percentage or the old flat rate
Perplexity	Free-tier Research allowance cut from 5/day to 1/month, corrected from OneHuman's own earlier miss	High	Track both free-tier clocks separately — Pro Search resets daily, Research resets monthly
Anthropic	Sonnet 5 $2/$10 pricing made permanent (firm, dated) vs. Claude Code's 4th limit-boost extension (hope, no date past Aug 31)	High	Plan Claude Code usage as if the boost could revert after Aug 31; treat Sonnet 5 API pricing as settled
Gemini	Named in Pebblous's model-verification scorecard via a gemini-cli bug (developer tool only, not the consumer app)	Low	No action for consumer Gemini app users; developers on gemini-cli should request explicit model IDs, not generic -flash version strings
Grok	Grok 4.6 launched Aug 12 — new flagship, 500K context, API pricing unchanged	Low	No significant changes for consumers
Copilot	GitHub Copilot's free-tier limits and included models publicly documented for the first time this cycle	Low	No significant changes for consumers
Mistral	No significant changes	—	—

---

The Watchdog Call

August's real pattern is that disclosure and accuracy are not the same thing, and four vendors demonstrated it without needing to coordinate. A company can tell the truth in a press-releasable number and still leave the reader with the wrong impression of what they'll pay — DeepSeek's 1,114% and OpenAI's 80% are both real, verified figures that pointed away from the number that mattered. Anthropic's split between a dated pricing commitment and an undated usage hope shows the same instinct applied to promises instead of numbers: say something reassuring, keep the actual commitment optional. None of this required lying. It required knowing which number or which word would get repeated, and picking that one. Expect September to bring either a fifth Claude Code extension landing close to its deadline — consistent with the pattern in all four before it — or the first sign of whether "we hope" ever converts into something dated.

The instruments that caught this — a 15-day pricing sweep run independently of vendor announcements, and continuous monitoring of primary sources rather than press coverage of them — are why this report exists in the form it does. They are not the story. The gap between the loudest number and the real one is.

---

Verified by OneHuman · August 31, 2026

---

Independent. AI-assisted. Human-verified. No ads. No affiliates. No investors.