There is evidence that Claude is gaining ground. There is much less public evidence for the stronger story that OpenAI's overall usage and revenue collapsed over the last two weeks because its models became worse.
As of October 8, 2026, the useful distinction is between website traffic, coding-agent activity, paid revenue, and the quality of an individual task. Those measures can move in different directions. Here is what the available sources establish—and what should actually change your choice of coding agent.
Claude's web growth is measurable
In a September 24 analysis, Similarweb reported that Claude's share of visits among the major AI chatbot websites rose from 2% to 9.6% in the twelve months through August 2026. ChatGPT's share fell from 78% to 57%.
That is evidence of a shift in consumer website attention over that period. It is not a measurement of Codex CLI sessions, API tokens, or the latest two weeks. Similarweb explicitly says its data cannot directly establish API consumption or company financial performance.
A falling share also needs a denominator: share measures a product relative to the group being tracked. It cannot, by itself, tell you how many paying customers left or why any particular user changed tools.
What about OpenAI revenue?
Axios reported rising OpenAI annualized revenue on September 29.
That reporting is not an audited financial statement. It also does not settle whether a company met a particular internal forecast. A claim about missing expectations needs a named target, an actual result, the same accounting period, and comparable definitions. Annualized run rate, recognized quarterly revenue, and profit are different measures.
The sources reviewed for this article do not establish a recent OpenAI revenue collapse. We therefore cannot use that claim as a reason to buy a competing coding tool.
Do complaints prove that Codex is getting worse?
An individual failure can be real without establishing a population-wide decline. A report with the starting files, model, settings, failed output, and acceptance check is useful evidence about that run. A collection of complaints without a denominator does not reveal the percentage of successful and unsuccessful tasks.
Nor does worse output identify its cause. A changed instruction, context handling, tool failure, model configuration, service issue, or routing change can require different remedies. Claims of deliberately serving a lower-quality model need request-level evidence; asking the assistant which model it is does not supply that evidence.
Our earlier GPT-6 Astra quality-report review distinguishes acknowledged issues from unproven explanations. It retains its September reporting cutoff; it should not be read as a fresh investigation of every October incident. If your current problem is practical, use the Codex slowdown and switching guide.
Service changes and stock-market stories need their own evidence
AP reported the Sora app shutdown in March. That is a specific product change with a specific date. It does not establish that Codex was shut down or that every OpenAI service lost demand recently.
The same discipline applies to market claims. Private-company valuations, transaction prices, and a listed technology company's share price are different things. A model launch followed by a stock move does not isolate the cause; earnings, interest rates, positioning, and other news can occur at the same time.
This article does not derive a trading forecast from coding-agent sentiment. For a developer deciding what to use tomorrow, the missing link is more immediate: which available configuration completes the work reliably?
A stronger reason to try Claude Code
Try it because it might solve a task that is currently consuming your time. The Claude Code coding evidence provides a starting point, while your own repository supplies the test.
Take one recent task that required repeated corrections. Save an independent copy of its initial state, give the new agent the same requirement, and judge the result against the same acceptance check. Include your review time and the actual bill. You do not need a theory about a provider's revenue to recognize a better outcome.
If you want both tools in the project, Agent.Space provides Workspaces with persistent files and separate Sessions for supported Agents and compatible models. Use isolated copies for comparison, or a clear sequential handoff to continue existing work. A new Session does not automatically inherit the old conversation, and an independent product does not transfer your official subscription allowance.
The practical goal is a project that can keep moving as the tools change. Choose from verified task results and current availability, then revisit that choice when the evidence changes.
Reporting cutoff: October 8, 2026. Traffic figures are attributed estimates for the stated period. Agent.Space has not measured aggregate Codex adoption, provider revenue, or a causal effect on stock prices.