My personal tier listSaved
AI Coding Agents Tier List
- 12 personally ranked
- 0 not ranked yet
- 8 with receipts on file
My tier list updated 2026-10-23
My call, from using these agents on real work. Select any agent for its receipts.
- S
- A
- B
- COnly for narrow, well-fenced tasks.
- D
Not ranked yet
Dossier
My tier: S1 receipt on file
Claude Opus 5.5
Anthropic · Claude Code
Runs at max effort in my Claude Code sessions.
Dan's review
Claude Opus 5.5 is extremely advanced, above all other models currently in many areas. The all-rounder MVP.
My view from daily use.
My tier: A0 receipts on file
Claude Sonnet 5.5
Anthropic · Claude Code
In my agent rotation.
Receipts on file
No report published for this agent yet.
My tier: D0 receipts on file
Claude Haiku 5.5
Anthropic · Claude Code
In my agent rotation as a fast helper tier.
Receipts on file
No report published for this agent yet.
My tier: A4 receipts on file
DeepSeek V4.1 Flash
DeepSeek · Claude Code (DeepSeek API)
66,428 September 2026 requests used the deepseek-flash alias; the serving checkpoint was not verified.
Receipts on file
- Head-to-headOct 19, 2026Extra review cost more on both repairsDoes an extra AI review step pay off? A small DeepSeek trial said no
- Cost receiptOct 17, 2026About 1 cent per million tokensMy real DeepSeek bill: 12.31 billion tokens for $122.28
- Self-reportOct 14, 2026Self-reported, not measuredI asked 8 AI coding agents for their biggest weakness
- IncidentSep 14, 2026Upstream hang, not my setupDeepSeek Flash answered 200 OK and sent nothing for two hours
My tier: A1 receipt on file
GPT-6.1 Sol
OpenAI · Codex CLI
In my agent rotation.
My tier: S0 receipts on file
GPT-6 Astra
OpenAI · Codex CLI
In my agent rotation.
Dan's review
GPT-6 Astra is the strongest architect.
My view from daily use.
Receipts on file
No report published for this agent yet.
My tier: D0 receipts on file
GPT-6 Luna
OpenAI · Codex CLI
In my agent rotation.
Receipts on file
No report published for this agent yet.
My tier: S3 receipts on file
Gemini 4 Argon
Google · Gemini CLI (Vertex AI)
Also reviews finished MythREAL episodes.
Dan's review
Gemini 4 Argon is the industry-leading native-video model.
My view from daily use.
Receipts on file
- Field testOct 21, 2026Useful review, inconsistent gradesGemini 4 Argon reviewed a finished episode: 14 clips, one flag, one inconsistency
- IncidentOct 14, 2026A client timeout, not the modelGemini 4 Argon kept hanging in Gemini CLI. The cause was a 60-second timeout
- Self-reportOct 14, 2026Self-reported, not measuredI asked 8 AI coding agents for their biggest weakness
My tier: S1 receipt on file
GLM-5.3
Qoder CLI
In my agent rotation.
Dan's review
GLM-5.3 is the industry-leading cyber model.
My view from daily use.
My tier: A1 receipt on file
Sonus
Qoder CLI
In my agent rotation.
My tier: A1 receipt on file
Qwen 3.8 Max
Qoder CLI
In my agent rotation.
My tier: B1 receipt on file
Muse Wonder Full 1.0
Muse Code
In my agent rotation.
How to read my rankings and evidence
I rank these agents from daily use on my projects. The reports show specific costs, failures and results. My rankings are personal judgments; a controlled comparison needs the same task and conditions for every agent.
Method for controlled comparisons
- Every agent in a round gets the same real task, the same repository state and the same time box.
- Every run is recorded: logs, cost, time, failures, and whether the result passed review.
- Each controlled round ends with a verdict per agent, supported by published evidence. These results are separate from my personal placements.
- Vendor benchmark tables, pricing pages and agents' self-reports are context, not my measured results.
- Major releases are reasons to revisit a comparison. A new model release does not by itself establish a new result.
Publication history
- Dan’s saved personal placements are shown separately from controlled comparisons, with his reviews and available reports.
- The comparison method and evidence roster were published; no controlled head-to-head standings were established.
Independent testing. Sponsorships are always disclosed, and verdicts are never for sale.
Full testing method