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 intelligence 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
Orchestration at startling scale. Users observed five-level nested subagents and 105 agents in a single session. The new agent view turned the terminal into a control plane.
Accuracy can pay for itself. One practitioner attributed a 40% cost reduction to improved accuracy:
“And now my costs are actually down 40%, presumably because of the better accuracy.”
CLAUDE CODE DISCORD USER · JUL 6
Release velocity outran reliability. Malformed tool calls, retry loops, timeout ceilings and ignored project rules repeatedly made interactive work unusable.
Long sessions lost the plot. Reports described circular reasoning, compaction loss, runaway subagents and unauthorized actions.
Bound the agents, tokens, permissions and behavioral drift—without neutering the system.
02 · THE MOMENTUM LEADER
It finished work that rivals drifted away from. Users reported plans drifting in Claude while Codex completed the implementation in one shot.
The $20 plan went further than expected. Code-review users praised allowance value; bankable limit resets felt like an effective price cut.
Windows and desktop state were brittle. Missing sessions, mixed queues, broken model pickers and indefinite “Working…” states made production use risky.
The context ceiling felt artificial.
“That’s basically a deal breaker for my research flow—one question often uses 200–300k tokens.”
DISCORD USER · JUL 9
Keep tools, task state and history intact across sessions, compaction, platforms and updates.
03 · THE HARNESS BECOMES A PLATFORM
Cloud agents became operational. Reusable VMs, visible terminals, `/in-cloud`, parallel branches, local handoff and PR babysitting moved beyond remote chat.
The context stack handled real scale.
“Subagents have been great… Mine is over 100k LoC and connected to three other services.”
R/CURSOR USER · JUL 8
Composer’s speed sometimes displaced reasoning. Users saw lazy automation on tasks that required deliberate implementation.
Headless did not equal IDE. The same task took about 3m15s in the IDE and 26 minutes through the SDK—with no semantic searches and 122 edit calls.
Translate models, cache, fast mode and tokens into productive hours and a believable bill.
04 · DISTRIBUTION MEETS AGENT MODE
A non-code workflow became an application. Cassidy Williams dragged a Photoshop template into plan mode, described the workflow and generated a functioning desktop tool.
“For this small little problem, I can now just click a button.”
CASSIDY WILLIAMS · MAY 19
Copilot became a multi-model surface. Local models, automatic routing and model swarms pushed it well past completion.
June pricing damaged trust. Reports described roughly 3× renewal increases and a shift toward less predictable consumption billing.
The agent workflow remained fragmented. Plan/agent switching was cumbersome; shared CLI tools initially raised review cost while finding fewer issues.
Show users what routing chose, what a task will cost and why agent mode is better than completion.
05 · THE BEAUTIFUL WILDCARD
Frontend aesthetics looked designed, not generated. Practitioners praised its layouts, visual judgment and escape from default “AI UI.”
A real niche win over Codex.
“Gemini is beating Codex hands down by a huge margin.”
DISCORD USER, SMALL GRAPHICAL APPS · JUL 1
Beautiful pages contained invented facts. One evaluation praised the visual result while finding nonexistent integrations and fabricated workflow claims.
Quota reality did not match the meter. A single-page project consumed a large share of a weekly allowance and still hit quota despite visible capacity.
“The design… looks fantastic, but the information is hallucinated in a lot of different places.”
COLE MEDIN · MAY 22
Preserve the visual imagination; remove fabricated repository facts and architectural assumptions.
BASE CASE · THROUGH APRIL 2027
This forecast concerns practitioner preference and serious active usage—not bundled seats or revenue. Product launches can overturn it; current trajectory sets the prior.
Switcher momentum, value and less drift. Fixable product debt.
Harness, cloud execution and workflow data compound.
Enterprise distribution outpaces developer preference.
Specialist UI adoption; generalist trust remains weak.
Still elite; reliability threatens relative share.
The market will not be won by the highest benchmark. It will be won by the first system to combine intelligence with durable memory, predictable cost, observable execution, safe rollback and stable behavior.
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 · 119 observations
Problem defeated: false completion · 41 observations
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.comHOW THIS ISSUE WAS MADE
The evidence window runs from May 18 through July 13, 2026. TruthAPI searches covered Reddit, Discord, GitHub issues and releases, YouTube, blogs and Twitter. Broad scans identified the market cohort; focused scans separated praise, surprise, regressions, cost, context and reliability; full records verified the quotations below. Cross-source waves were used to locate inflection points.
C1 agent view · C2 nested subagents · C3 40% cost report · C4 malformed tool calls · C5 timeout regression · C6 quality regression
O1 usability comparison · O2 lower-drift one-shot · O3 $20 plan value · O4 banked resets · O5 Windows regressions · O6 context deal-breaker · O7 indefinite working state
U1 cloud-agent release · U2 100k-LoC workflow · U3 config awareness · U4 Windows performance fix · U5 speed/reasoning tradeoff · U6 headless slowdown · U7 plan-value concern
P1 plan-mode application · P2 model swarm · P3 local models · P4 redesigned CLI · P5 pricing change · P6 plan/agent UX · P7 code-review regression
G1 graphical-app comparison · G2 full-stack/UI praise · G3 visual quality · G4 hallucination/quota evaluation · G5 crash/state loss · G6 Antigravity downgrade · G7 rate-limit latency
These records are directional qualitative evidence, not a representative survey. The “top five” reflect visibility, not audited share. Model sentiment can be conflated with harness quality. Forecast probabilities express editorial confidence, not measured frequencies.
The exact TruthAPI queries, selected record IDs and processed conclusions are preserved in findings/ai-coding-systems-truthapi-research-log.md.
First edition · Produced by TruthAPI intelligence · gettruthapi.com