THE TOKEN REPORT
Eight weeks of user evidence across Claude Code, Codex, Cursor, GitHub Copilot and Gemini reveal a new competitive axis: not who writes the best code—but who can be trusted with the longest leash.
What thousands of builders already learned. Inside your agent’s reasoning loop. Updated daily.
THE BIG SHIFT
The last eight weeks were not a simple model horse race. Every major coding system gained power. The surprise was what users started complaining about once that power arrived.
Cloud agents, nested subagents, automatic routing and million-token models expanded what developers could delegate. But the conversation moved just as quickly toward context loss, runaway usage, invisible quotas, session failures and output that looked right while being wrong.
The durable advantage is shifting from model capability to operational trust: durable state, observable execution, predictable cost, stable releases and safe rollback.
“Users cannot reliably predict the outcome, cost or collateral behavior of a long-running coding-agent task.”
THE TOKEN REPORT, BASE-CASE THESIS
01 · CURRENT LEADER, HIGHEST VOLATILITY
The harness is the advantage. TruthAPI connects Claude Code to subagents, skills, MCP, plan-execute workflows, memory and long-horizon editing—not merely a strong model.
Practitioners still notice the difference.
“Claude Code is best because it has dynamic workflows.”
CLAUDE CODE DISCORD USER · JUL 10
Failure can look like success. A fresh issue reports a clean code review after subagents had failed at the spend limit—while billing continued.
Scale magnifies context loss. One current task spawned four subagents; eight compactions per agent produced 40 reported off-track incidents. A new rate-limit issue adds service risk.
Bound the agents, tokens, permissions and behavioral drift—without neutering the system.
02 · THE MOMENTUM LEADER
The switcher signal repeats. Moves from Claude to Codex appeared in both fixed evidence windows, alongside current claims that Codex is the better implementation engineer.
The model line moved again. GPT-5.6 shipped across Codex and the API, while practitioners compared OpenAI's usage resets favorably with Anthropic's.
The harness still trails. Current practitioner evidence says the underlying model is competitive while workflows and subagents remain less mature than Claude Code.
Product state is not durable yet. Fresh issues report Windows freezing when changing projects and a broken agent-thread picker.
Keep tools, task state and history intact across sessions, compaction, platforms and updates.
03 · THE HARNESS BECOMES A PLATFORM
Cursor is becoming portable. ACP exposes its agent inside Zed and IntelliJ instead of confining it to one editor.
Workflow data may become a moat. Grok 4.5 was described as trained with Cursor data from real developer workflows. Mobile and fast-edit preferences broaden the surface.
Cursor's edge is increasingly the workflow—not a captive model.
THE TOKEN REPORT · SYNTHESIS
Routing opacity erodes trust. A current report says Cursor switched from Grok to Sonnet without intent and consumed API usage.
Parallel agents lack shared state. Fresh evidence describes Claude Code and Cursor silently contradicting each other on the same repository.
Translate models, cache, fast mode and tokens into productive hours and a believable bill.
04 · DISTRIBUTION MEETS AGENT MODE
The agent now reaches the work queue. Copilot can accept a Jira ticket, stream progress, change code and open a pull request without leaving Jira.
GitHub owns the execution surface. Its cloud agent can research, plan, branch, test and iterate inside a GitHub Actions environment.
Advanced users still separate it from long-horizon agents. Current comparisons describe Copilot mostly as autocomplete and weaker on multi-step context.
A local win did not generalize. GitHub says the focused instructions that improved code review did not improve Copilot CLI's broader exploratory work.
Show users what routing chose, what a task will cost and why agent mode is better than completion.
05 · THE BEAUTIFUL WILDCARD
The technical surface is substantial. Gemini 3.5 Flash documents a 1M-token input window, code execution, search grounding and fast multi-step loops. Antigravity adds a managed sandbox.
A non-programmer shipped a working artifact.
Hundreds of prompts produced a playable browser game in one 2 MB HTML file.
R/GEMINIAI USER · JUL 15
Current users report fragile coding work. One migrated from Antigravity to Claude after repeated errors; another hit a limit after roughly 30 minutes and switched providers.
Approval boundaries regressed. A current report says Antigravity changed files without asking where Gemini CLI and Claude Code required approval.
Preserve the speed and visual imagination; remove fabricated repository facts and unsafe execution.
BASE CASE · THROUGH APRIL 2027
Relative practitioner traction among serious active users. Direction is qualitative; confidence is editorial—not a measured probability.
The deepest agent-workflow harness in the graph is also accumulating spend, compaction and reliability failures. The pattern appears in both fixed evidence windows.
Claude remains preferred for multi-file reasoning and dynamic workflows. Stability-first releases plus hard spend controls could erase the risk.
Switcher signals recur across both windows. GPT-5.6 adds capability while current practitioners describe Codex as a stronger implementation engineer.
The harness still trails Claude's workflows and subagents. New Windows and thread-state bugs could recreate the same trust problem.
Cursor is compounding a workflow and data layer: ACP portability, editor integration and model training informed by real developer work.
First-party harnesses can narrow the moat. Opaque model switching and usage charging could turn orchestration into a liability.
BASE CASE · THROUGH APRIL 2027
Copilot now reaches the issue queue, repository, Actions and pull request. Distribution should lift enterprise use; current advanced-user comparisons still favor other long-horizon agents.
If GitHub turns issue, code, CI, policy and review into one coherent loop, distribution becomes product advantage and power-user preference could rise sharply.
Fast loops, 1M-token context, code execution and a managed sandbox support prototyping. Current quota, approval and coding-reliability reports weaken the generalist case.
Google has speed, context, grounding and distribution. A unified developer surface with reliable approval and quotas could create broad gains quickly.
Issue 002 will score these calls with the same symmetric queries and consecutive fixed windows. Numeric probabilities return only after the calls earn calibration.
Capability is abundant. Durable state, bounded execution, predictable cost, observable work and safe rollback remain scarce.
DON’T PICK A RELIGION. DESIGN A PORTFOLIO.
Use Codex where session stability is proven; keep Claude for deep reasoning and orchestration with explicit agent and spend caps.
Best fit for codebase context, local-to-cloud handoff, background agents and pull-request flow. Track spend outside the product.
Exploit repository and policy integration. Benchmark routed models and agent mode against a simpler completion baseline.
Delegate UI exploration and graphical apps. Never accept product facts, APIs or architecture without independent verification.
WHAT BUILDERS ARE DOING NOW
The common use case is no mystery. The important signal is what happens when builders connect code execution to a measurable outcome.
Give the agent a concrete change and an objective check it can run.
Method: verification loop
Problem defeated: false completion
The agent proposes, runs, measures and revises—not merely writes the research code.
An Auto Research workflow found word-error-rate improvements beyond conventional hyperparameter optimization.
Frontier coding agents reproduced computational findings as reliable workflow executors.
Signal: strongest for Claude Code, with additional Codex research evidence.
Pattern: these differ by system. The interface determines the edge.
The product is no longer always software. Sometimes software is simply how the agent reaches the outcome.
Explore what builders are learning → gettruthapi.comSEE THE GAP. CLOSE THE GAP.
Research-normalized ranges. “Elite” is a target—not a measured percentile. Count only verified work.
START HERE: Separate implementation from verification. It improves judgment before you change a model, prompt or budget.
WHAT THE EVIDENCE CHANGES
Better models help. Better operating systems for agents compound the gain.
FastContext reduced token use while improving issue resolution by 5.5%.
arXiv:2606.14066Co-Coder reached 2.10× speed and 35% lower API cost at the same time.
arXiv:2606.00953ScarfBench found only one fully behaviorally equivalent result in 204 tasks.
arXiv:2605.06754MPAC cut coordination overhead by 95% in controlled multi-agent review.
arXiv:2604.09744That is the opportunity shown on the previous page.
HOW THIS ISSUE WAS MADE
The current evidence window runs from May 25 through July 20, 2026; the comparison window is March 30 through May 24. All five products received the same progress, outcomes, friction and competition scans, then a prior-window scan, a graph traversal and full-record inspection. Waves were used only as a cross-check.
One exact-ID-deduplicated TruthAPI record: an event, an object or a documentation passage. It is not a unique person, company, vote or controlled observation.
First-party documentation, direct artifacts, GitHub issues, reproducible outcomes, repeated practitioner reports and signals that survived both fixed windows.
Raw mention volume, a single viral post, bundled seats, product marketing or rank alone. Counts describe retrieval coverage—not adoption or market share.
Each product received 1,200 current result slots and 300 prior-window slots. Exact IDs were deduplicated before candidate selection. `getSet` inspected the decisive records in full.
This is directional qualitative evidence, not a representative survey. Source communities overlap. User reports can conflate model and harness quality. Discord and Reddit signals are self-reported. Direction bands express editorial judgment.
Issue 002 will score every call using the same protocol. Numeric probabilities return only after repeated forecasts earn calibration.
The 39 queries, full return payloads, selected records, analysis and every reply time are preserved in research/token-report-refresh-2026-07-20/.
WHAT WE CANVASSED AND ANALYZED
Broad retrieval found the field. Graph traversal connected each product to its methods and problems. Full records tested the decisive claims.
Repeated evidence over isolated heat. Direct reports, artifacts, first-party docs and research carried more weight.
Contradictions stayed visible. Every forecast includes a condition that could reverse it.
Denominators stayed honest. Record counts measure coverage—not people, prevalence or audited market share.
DON’T TAKE OUR WORD FOR IT
Ask the question you really care about. Put it into TruthAPI and inspect what builders, repositories, research and product changes say now.
Show the strongest current evidence that contradicts Claude Code’s agentic-workflow lead.
Compare Claude implementation + Codex review with the reverse workflow. Where does each fail?
Find the strongest real-world case for GitHub Copilot among advanced practitioners—not bundled seats.
What advantages does Cursor have beyond model choice? Separate the harness from the underlying model.
Where does Gemini beat Claude or Codex today? Return concrete tasks, artifacts and counterevidence.
Which conclusions in this report have changed since July 20, 2026—and what new evidence changed them?
Current evidence. Connected context. Returned inside your agent’s reasoning loop.
PROBE THE EVIDENCE → GETTRUTHAPI.COMThe research archive preserves all 39 query arguments, full return payloads, exact reply times, deduplication metrics, selected records and editorial analysis. Counts describe returned evidence—not unique people, audited market share or controlled trials.