Codex Rewrites Its Own Prompt Every Turn, Claude Code's Boost Runs to August, NVIDIA's Benchmark Swings Wildly

A developer running OpenAI's Codex CLI against Amazon Bedrock filed a GitHub issue this week with a number that's hard to ignore: over four days of production use, August 5 to 8, cache writes ate up roughly 85% of the model's estimated spend on GPT-5.6 Sol, per a Cost Explorer-derived estimate in the report. One session logged 76 requests, 6.709 million cache-write input tokens, and zero cached input tokens -- an average of about 88,000 cache-write tokens on every single call. CloudWatch showed no client errors in that window, so this wasn't a broken integration retrying failed requests. Everything succeeded. It just succeeded at full write price, every time.

That's not a subtle inefficiency. It's close to a total failure of caching, in a category of tool -- agentic coding assistants -- that depends on caching to be affordable at all.

Here's the mechanism. Tools like Codex CLI resend a long, largely unchanging prefix on every turn -- system instructions, tool definitions, prior context -- followed by whatever actually changed. Prompt caching exists so a provider doesn't reprocess that stable prefix from scratch each time; it stores it once and charges a fraction of the price on subsequent reads. AWS documents an explicit cache mode for GPT-5.6 on Bedrock built for exactly this kind of workload: long stable prefix, changing tail.

The catch: native Codex CLI requests to Bedrock's Mantle Responses API don't expose the fields needed to opt into that mode. Codex already generates a session-scoped prompt_cache_key, but neither its HTTP nor WebSocket request types include the prompt_cache_options or a cache-breakpoint field that would let it claim a cache read. Bedrock's built-in provider config only handles transport and authentication, not request-body structure, so there's no config-level workaround either. Codex ends up paying full write price for the same stable prefix, turn after turn, instead of the cheaper read price after the first request. The filer notes this builds on an earlier report, #35300, but adds independent production evidence from the native Bedrock provider specifically.

The filer is careful not to overclaim: cold starts, genuinely new prompts, session forks, and context compaction can all legitimately require fresh writes, and the numbers are usage-derived estimates, not finalized AWS invoices. But 85% of spend and an 88K-token average, both drawn from the filer's own Cost Explorer and CloudWatch data, look less like occasional cold starts and more like a missing plumbing feature. The proposed fix is narrow: serialize prompt_cache_options, add a typed cache-breakpoint field, gate it by provider and model capability, and surface cache reads versus writes in per-turn usage telemetry -- so the next person hitting this sees it in the numbers instead of finding out from a bill.

Anthropic has extended its limited-time promotion boosting weekly usage limits in Claude Code by 50%. The support page says it plainly: "We've extended this promotion." The window now runs from May 13 through August 31, 2026 at 11:59 PM PT — roughly three and a half months of extra weekly capacity for developers who hit the ceiling before the week is over. That it needed extending at all is the first clue: a promotion Anthropic could have let quietly expire instead got a new end date, which suggests the original one was working well enough to keep running.

The mechanics are simple. Anyone on a Pro, Max, Team, or legacy seat-based Enterprise plan gets 50% more weekly usage in Claude Code, automatically, with no opt-in required. It applies everywhere the product lives — CLI, IDE extensions, desktop, and web — and developers can check their new ceiling by running /usage in the terminal. Free plans and consumption-based Enterprise seats are excluded, which draws a clean line between subscribers and the metered usage Anthropic already bills by the token. The offer has no cash value, is not transferable, and can't be combined with other promotions. After August 31, 2026, weekly limits return to standard levels, with no billing or plan changes on either end — so nobody needs to remember to cancel anything or adjust a seat.

One boundary is worth noting: the promotion is scoped to Claude Code only. Weekly limits for Claude on web, desktop, and mobile, and for Claude Cowork, are unchanged. That narrow focus is the tell — Anthropic isn't handing out broader usage relief across its product line; it's releasing pressure at the exact point where agentic coding runs get throttled mid-week.

What the boost doesn't touch is just as telling. The 5-hour session limit is untouched — only the weekly ceiling moves. A single coding session was never really the bottleneck for developers running Claude Code through multi-step refactors and repo-wide migrations across several days; the wall people actually hit is the weekly one, when earlier exploratory work in the same week has already spent the budget. Loosening that specific dial, and leaving the session-level one alone, says something precise about where the real friction has been: not in any single burst of work, but in the cumulative total across a week.

It also says something about internal accounting. Extending a 50% boost means absorbing extra inference cost for months across plans that, unlike consumption-based Enterprise, aren't metered per token. That is not a call a company makes lightly if usage were spiking toward margin-threatening territory. The more plausible read: demand from agentic coding workflows is still below whatever ceiling Anthropic modeled when it set standard weekly limits, and loosening the cap costs less than watching developers hit a wall mid-week and reach for a competitor's coding CLI instead.

NVIDIA just published the numbers behind its own "verified skills" program, and the honest answer is: it depends enormously on which skill you install. The company open-sourced SkillEvaluator, a tool that runs the same coding task twice — once with a skill loaded into Claude Code, Codex, or Cursor, once without — then scores the difference. Across a snapshot of 300-plus verified skills spanning 30-plus NVIDIA products, taken August 12, 2026, NVIDIA reports an average 31-point lift on a 100-point scale, or 39 points excluding Security, which already scored 97 out of 100 with no skill installed.

That average is the headline. The spread underneath it is the real story. NVIDIA's own token-tracking data shows one skill, jetson-optimize-memory, cutting token usage on a single run from 617,306 to 142,540 — a 76.9% reduction — while trimming execution time from 474.9 to 220.0 seconds, a 53.7% cut. A different skill, cuopt-install, did the opposite in NVIDIA's own single-attempt test: tokens rose from 25,227 to 55,582, a 120.3% increase, with execution time climbing from 34.0 to 41.1 seconds, up 20.8%. Same method, same "verified" label, opposite outcome — these are two single-attempt examples, not catalog-wide averages, but they sit in the same writeup as the headline number, which on its own implies uniform improvement.

The deeper finding: harness choice barely matters next to which product and task is being tested. Claude Code and Codex differed by only about 5 points on average, but per-product Skill Lift ranged from roughly +2 to +46 depending on the domain and how the evaluation dataset was written. Broken out by dimension, Correctness climbed from a 46 baseline to 87, and Effectiveness from 39 to 78 — gains of 41 and 39 points. NVIDIA also flagged something less flattering to the category: Discoverability and Efficiency, whether an agent even loads the right skill and avoids wasted steps, started at baselines of just 42 and 43 out of 100 with no skill installed, climbing by 40 and 35 points once a skill was in place. A skill that goes undiscovered, or loads for irrelevant tasks, doesn't just fail to help — every skill in an agent's context competes for its attention, so the wrong one can drag performance down.

That's a useful admission from a vendor shipping a marketplace of these things, and it means the "verified" badge does less work than it sounds like. NVIDIA's Tier 1 and Tier 2 checks only confirm a skill is safe and non-duplicative before it ships; only the live Tier 3 run, on real tasks inside an isolated sandbox, tells you whether the skill actually helps.

A security audit found more than 21,000 Model Context Protocol servers sitting open on the public internet — the connective tissue that lets AI agents call tools, run shell commands, and touch credentials on a user's machine. Among a 640-server sample, 91.8% had no OAuth configured at all, leaving those tool endpoints exposed to remote command execution. That same report counted 687 individual instances showing the identical gap. MCP has shipped inside tooling tied to roughly 150 million downstream downloads, according to that report, so a missing-auth default here isn't a lab curiosity — it's sitting behind agents already wired into real developer workflows, each one a direct line to local shell commands and whatever credentials that agent already holds.

According to Artificial Analysis, Alibaba's 27-billion-parameter Qwen 3.8 posted a 52 on its Intelligence Index. That ties GPT-5.6 Luna (max), a system whose scale Alibaba doesn't need to match. Two much larger models edged past it on the same index: one with 753 billion parameters and one with 1.7 trillion, both scoring 53. The entire range, from a 27-billion-parameter model up to one more than sixty times larger, is one point wide. That's the kind of gap that used to separate a frontier model from last year's mid-tier release — now, on this index, it separates the smallest model on the board from the largest.

According to OpenAI, Asana used its Codex agentic coding tool to close out a backlog of legacy code migrations and technical debt the company had been putting off, and OpenAI says the work — compressed into about two weeks — is what Asana estimates would have taken five years of engineer time. Both numbers come from OpenAI's own case study, not an independent audit, but the shape of the claim is telling: this isn't autocomplete finishing a function, it's agents grinding through the unglamorous, backlogged migration work that engineering teams typically deprioritize indefinitely because no one wants to spend a sprint on it. If the self-reported figures hold up under scrutiny, this is the clearest vendor proof yet that agentic coding can absorb bulk refactoring, not just autocomplete.

According to an analysis by Tipiirai, Bun's maintainers are rewriting the JavaScript runtime's core from Zig to Rust using autonomous Claude agent loops running around the clock. Those agents have opened roughly 15,800 pull requests against the repo -- only 790 merged, about 1,600 closed without merging, and more than 5,000 still sitting open, with agent accounts hitting as many as 1,000 pull requests in a single day, far outpacing what any review team could triage. That gap points to something worth remembering about agentic coding: it scales output, not judgment, and without a merge gate the PR count just keeps climbing past anyone's ability to confirm the code is worth keeping.