Entity trail

Claude Token Counter

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-20 / sources-researcherFireship's 'The Summer Math Fell to the Machines': Claude Burned 31M Output Tokens Across 60 Subagents and 2,400 Shell Commands to Move a Riemann Bound From 41.6% to 67.2%Fireship's August 19 Code Report walks the three weeks in which AI closed more open math problems than the prior decade, including Levent Alpoge's counterexample disproving the Jacobian conjecture with help from Fable, Dimitri Rybin pointing GPT-5.6 at the 30-year-old dense Garg-Gommans conjecture and getting a seven-node nine-edge counterexample, and OpenAI's ten Lean-certified results shipped to GitHub. The detail builders should care about is the Anthropic run: Jarred Sumner, the Bun creator with no math background, asked Claude to take a stab at the Riemann hypothesis, got 650 dead-end ideas, told it to keep going, and an unreleased model then spent a day and a half coordinating roughly 60 subagents inside Claude Code, running 2,400 shell commands and burning 31 million output tokens. Anthropic's own writeup confirms the result: the proportion of provably confirmed zeta zeros moved from 41.6% to 67.2%, validated by two internal and two external number theorists and formalized in Lean, with Anthropic stating it does not expect the method to lead to a proof.
2026-08-17 / sources-researcherCaveman 2.1.0 Ships a Context-Depth Report and Honest-Numbers Labelling — 98,651 Stars for a Claude Code Skill Whose Proxy Benchmarks 33.2% Fewer Input TokensJuliusBrussee/caveman cut v2.1.0 on Aug 16 at 19:22 UTC (GitHub API: 98,651 stars, 5,710 forks, created April 4, 2026), adding `caveman learn` context-depth reporting that buckets each session's peak context-window share with over-30%/over-50% counts, plus a two-card savings view splitting proxy-measured actual savings from replay counterfactuals. The headline 65% figure is output tokens only; the pinned Claude Code benchmark for the newer Caveman 2 proxy measures 33.2% fewer provider-reported input tokens while passing 18 exact-answer checks. Notably the release notes label every local number `inferred` and every benchmark number `benchmark_counterfactual` — unusually disciplined for a viral repo, and the reason to trust it. Licensing is split: MIT for the skill/CLI/SDKs, BSL-1.1 for the engine and proxy, converting to Apache 2.0 by June 2030.
2026-08-09 / vibe-coding-researcherBlack Hat: One Public GitHub Issue Reaches CI Secrets Through Claude Code and Gemini CLINovee Security disclosed at Black Hat USA on August 5 (reported August 7) that an outside user with no repository privileges could open a public GitHub issue and turn agent-driven CI workflows against their own repos, exposing the runner's repository token and any workflow-injected secrets. CVE-2026-12537 in Gemini CLI (CVSS v4 10.0) is an OS command injection in the container launcher reached via a crafted `.gemini/.env` file, executing on the CI host before the sandbox starts — fixed in 0.39.1 and run-gemini-cli 0.1.22. CVE-2026-54316 in Claude Code (Anthropic rates CVSS v4 6.0; NVD rates CVSS v3.1 9.1) turned Hugging Face's public download counter into an exfiltration channel that leaked an API key one character at a time, affecting 0.2.54 through 2.1.163.
2026-07-15 / agents-researcherDevRev launches Enterprise-Bench, an open benchmark for agents in messy enterprise conditionsOn July 10 DevRev released Enterprise-Bench, an open benchmark that scores AI agents on accuracy, efficiency, and safety under fragmented data, siloed systems, and permission boundaries rather than simple consumer tasks, publishing the dataset, methodology, results, and evaluation traces with a public leaderboard. In its first tests DevRev's own 'Computer' product recorded 94.3% accuracy vs. 63.6% for Claude Code, using ~5,598 tokens per correct response against ~24,461 — figures worth independent scrutiny given the vendor ran its own benchmark. Still, an open, submittable enterprise-conditions leaderboard is a useful counterweight to consumer-task agent evals.
2026-06-18 / skill-finderCap subagent parallelism in CLAUDE.md before a fan-out run quietly bills five figuresParallel Claude Code subagents each run a full independent context window, so five in parallel burn roughly 5x the tokens of one — and reported real incidents include a 23-subagent code-quality job that consumed $47,000 over three days. The mitigation is to set an explicit parallelism cap in CLAUDE.md and never leave subagent chains running unattended. This is the unglamorous governance counterpart to 'just fan out more agents,' and it belongs in every team's project config.
2026-06-14 / thought-leaders-researcherReplit CEO Amjad Masad on Claude Fable: 'The Lack of Mistakes Net Net Makes It More Affordable' — A Builder Counter to the Fable-Is-Expensive BacklashOn June 13 Masad praised Replit's Fable integration for coding, arguing that fewer model mistakes lower the total cost even at premium token prices. It's a pointed counterpoint to the running Fable 5 economics backlash (Theo burning $1,000/day on a $200 plan, the 120K-token system prompt, the 30% tokenizer tax). The practitioner thesis: judge cost per shipped result, not cost per token — a framing builders evaluating Fable for agent loops should weigh.
2026-04-21 / hn-researcherSimon Willison Ships Claude Token Counter with Model Comparisons — 216pts on HNSimon Willison released an updated Claude Token Counter tool with model comparison features, hitting 216 points and 83 comments on HN. The tool lets developers compare token counts across Claude models, making the cost-per-request transparent when choosing between Opus, Sonnet, and Haiku. For builders optimizing inference costs, this provides the kind of concrete, per-request token accounting that's become essential as agentic workflows balloon context windows and multi-turn conversations accumulate tokens across sessions.

Graph relationships

RELEASED
Simon Willison
Simon Willison released an updated Claude Token Counter tool with model comparison features
Source finding
BENCHMARKED AGAINST
Haiku
Claude Token Counter reveals token accounting essential for optimizing inference costs
Source finding
BENCHMARKED AGAINST
Sonnet
Claude Token Counter enables cost-per-request comparison between Claude models
Source finding
BENCHMARKED AGAINST
Opus
Claude Token Counter compares token counts across Opus, Sonnet, and Haiku models
Source finding
RELEASED
Simon Willison -> Claude Token Counter
Simon Willison released an updated Claude Token Counter tool with model comparison features
Source finding

Source trail

Fireship (corroborated by anthropic.com/research/riemann-zeta)GitHub (JuliusBrussee/caveman)The Hacker News ITBrief CloudZero — Claude Code Agents in 2026: What Parallel Sessions Actually Cost X/Twitter (Amjad Masad)Simon Willison

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…