Edition 2026-09 · gpt-6
What should you actually use for agentic coding?
An evidence review of coding models, agent harnesses, costs, privacy, and local deployment — with recommendations for concrete tasks and operating constraints.
The state of play
Edition 2026-09
September’s releases justify new workflow trials. Astra/Codex and Fable 5.1/Claude Code form a shortlist for difficult work; incomplete benchmark conditions prevent a confident winner. GLM-5.3 deserves a fresh budget evaluation, with changed routing, mandatory reasoning, and model-specific license conditions. Benchmark evidence, GLM migration.
Keep a proven everyday workflow until another improves accepted outcomes or total cost. Fable 5.1’s cheaper cache reads can lower repeated-context bills, while its retention and migration requirements can change the approved route. No-training policies do not imply zero retention. Fable specifications, access and retention.
The local hardware options are untested pilot candidates. Productivity studies concern older tools; selected developer accounts do not establish consensus. Human review and bounded authority remain prerequisites. This edition reports source evidence and conditional recommendations, with no original comparative trials.
Snapshot: September 8, 2026, America/New_York.
14
models represented in the edition's setups
12
harness and execution-route labels represented
12
scoped privacy and training policy records
7
use-case recommendations for individuals & teams
Source window 2025-09-29 → 2026-09-08. Methodology v1.1.0. As of 2026-09-08. Changes since the baseline →
If you are X, use Y
The picker
Type your situation — budget, privacy, repo size, all-day volume — to filter the edition's recommendations and their conditions.
Daily implementation (Individual)
→
Keep Sol/Codex; reassessRequires: Tests; review. Alternative: Opus 5/Claude Code. Tradeoffs: Promotional rates; task costs unmeasured. Confidence: medium — Documented rates; limited experience. Reconsider if: Another setup reduces accepted-task cost. Claims: CR13;CR07;CR16;CH17;CO12
Hard debugging/refactoring (Experienced developer)
→
Trial Astra/Codex localRequires: Repro case; approved data route. Alternative: Fable 5.1/Claude Code. Tradeoffs: No clear leader; Fable retention/migration limits. Confidence: medium — Relevant system evidence; incomplete conditions. Reconsider if: Repo trials favor alternative. Claims: CR01;CR02;CR03;CR04;CR05;CH17;CO04;CO05
Bounded volume (Cost-sensitive developer)
→
Pilot GLM-5.3/OpenCodeRequires: Tests; review; approved route/protocol. Alternative: DeepSeek V4-Pro; GLM Flash. Tradeoffs: Aliases reroute; Coding Plan privacy unresolved; weight license separate. Confidence: low — Rates supported; plan policy unresolved. Reconsider if: Retries/review erase savings. Claims: CGL01;CGL03;CGL04;CGL06;CGL20;CR12;CR17;CO06;CGLP01;CGLP02;CGLP03
Greenfield/UI (Product developer)
→
Pilot Gemini 3.8/AntigravityRequires: Browser checks; acceptance criteria. Alternative: Existing Sol/Opus workflow. Tradeoffs: Managed-agent billing unverified; vendor demonstrations. Confidence: low — Comparative outcomes missing. Reconsider if: Measured review and defect costs improve. Claims: CR08;CR09
Bounded local tasks (Offline developer)
→
Pilot Qwen3.8-27B/4-bitRequires: Compatible quant/server; isolated tools; tests/review. Alternative: Smaller Qwen3.5-9B. Tradeoffs: 24GB fit/context/latency untested. Confidence: low — Weight arithmetic only. Reconsider if: Hardware tests fail thresholds. Claims: CGL09;CGL17;CGL18;CGL21
Controlled rollout (Existing GitHub team)
→
Trial Copilot Business/AstraRequires: Approve retention; cap spend. Alternative: Approved Codex Business route. Tradeoffs: Service retention unknown; staged access. Confidence: medium — Documented controls; quality unmeasured. Reconsider if: Policy or task trial fails. Claims: CH02;CH04;CH07;CH12;CH17;CH20;CH21
Unattended production (Engineering lead)
→
Insufficient evidence for automatic approvalRequires: Owner; sandbox; rollback. Alternative: Bounded background tasks; human approval. Tradeoffs: PR approval preview grants authority. Confidence: low — No reliable production success rate. Reconsider if: Audited trials satisfy acceptance criteria. Claims: CH03;CH20;CO07;CO08
No match — try a broader term, or see the full recommendations.
Models and harnesses
The contenders
Read each assessment with its task, configuration, and evidence limits. Scores from different evaluation setups do not establish a model ranking.
| Model | Harness | Configuration | Task | Strengths | Limitations | Claims |
|---|---|---|---|---|---|---|
| gpt-6-astra | Codex CLI/IDE local | Start high; escalate effort only after checking failure | Complex implementation and debugging | Steerable long-running candidate | No model-only superiority established; costly cache/long requests | CR01;CR02;CH17;CH20 |
| claude-fable-5-1 | Claude Code | High default; preserve prefixes and verify fallback route | Hard refactoring and long sessions | Cache-heavy escalation candidate | Retention eligibility; API breaking changes; human review needed | CR03;CR04;CR05;CR06;CH05;CH06 |
| claude-opus-5 | Claude Code | High; bounded task and tests | Everyday implementation | Lower list rates than Fable; vendor recommended starting model | Vendor guidance; no measured repo-specific preference | CR07;CR16 |
| gpt-5.6-sol | Codex | Medium starting effort; bounded task | Everyday implementation | Lower current token rates | Promotion; no accepted-task savings established | CR13;CH17 |
| gemini-3.8-flash | Antigravity managed agent | Medium default; allow billed reasoning | Greenfield/UI and bounded budget evaluation | Native managed route; managed-agent economics unmeasured | May spend more tokens; no independent advantage established here | CR08;CR09 |
| deepseek-v4-pro | Codex | Vendor Responses configuration; high starting effort | Budget workflow compatibility evaluation | Documented Responses integration | Not tested; Responses protocol has unsupported fields and tools | CR11;CR12;CR18 |
| glm-5.3 | OpenCode | Reasoning enabled; explicitly select effort; retain task tests | Budget-conscious implementation and debugging with human review | Low posted token rates; documented Coding Plan support | Task acceptance cost unmeasured; aliases reroute; account protocol restrictions; Coding Plan privacy mapping unresolved | CGL01;CGL04;CGL06;CGL20;CGLP03 |
| glm-5.3 | Claude Code | max effort; 400K context; 128K output; temperature=1; top_p=1; Tool Search disabled | Vendor Terminal-Bench 3.0 evaluation | Detailed vendor evaluation conditions | No independent reproduction or matched deployment measurements | CGL02 |
| glm-5.3-flash | OpenCode | Explicit effort; max default; multimodal model | Low-price, bounded implementation tasks to pilot | Lower durable list rates; multimodal input; MIT weights | Quality and accepted-task savings unverified; not a small local model | CGL05;CGL06;CGL08 |
| Qwen/Qwen3.5-4B | Qwen-Agent + vLLM/SGLang | Proposed 4-bit weights; KV precision and context to measure | 8 GB dedicated VRAM capacity-screening pilot | Small Apache 2.0 candidate with documented local agent route | Quant format, runtime memory and task reliability not tested | CGL17;CGL18;CGL21 |
| Qwen/Qwen3.5-9B | Qwen-Agent + vLLM/SGLang | Proposed 8-bit weights; KV precision and context to measure | 16 GB dedicated VRAM capacity-screening pilot | Larger small-model candidate with local MCP route | No measured agentic reliability or latency | CGL17;CGL18;CGL21 |
| Qwen/Qwen3.8-27B | vLLM/SGLang endpoint + chosen agent | Proposed 4-bit weights; preserve reasoning per model docs; client verification required | 24 GB dedicated VRAM local pilot | New dense model; Apache 2.0 designation; vendor coding improvement | Vendor benchmark does not validate this 4-bit setup; client interoperability untested | CGL09;CGL10;CGL18;CGL21 |
| Qwen/Qwen3.8-27B | Selected local serving stack and agent | Proposed 8-bit weights; engine/hardware support to verify | 64 GB unified-memory capacity-screening pilot | More weight precision headroom than the 24 GB tier | Unified memory shared with OS; neither engine compatibility nor latency measured | CGL09;CGL18;CGL21 |
| Qwen/Qwen3.8-27B | Terminus | Vendor benchmark; detailed configuration unknown | Vendor Terminal-Bench 2.1 result | Reported score of 73.0 | No 4-bit or commodity GPU inference established | CGL10 |
| kimi-k3 | Kimi Code CLI | Preserve full assistant messages; effort low/high/max; max default | Long-session engineering candidate | Native multimodality and 1M context documented | Vendor capability claims; custom license; huge self-hosting memory | CGL13;CGL14;CGL18 |
| MiniMax-M3 | Selected compatible agent | Choose enabled/adaptive/disabled thinking; integration test required | Multimodal coding candidate | Documented 1M context and serving routes | Custom license; no measured task economics or client validation in this review | CGL15;CGL16 |
| Qwen/Qwen3.8-Flash-Next | Selected compatible agent | Experimental preview; version/quantization and parser verification required | Larger-server experiment | Sparse architecture with offloadable embeddings | Custom AI Work Assistant/MaaS licensing; 6B active parameters do not determine total weight memory | CGL11;CGL12 |
| GPT-6 Astra | Codex | Max effort; Harbor evaluation using Terminal-Bench 4.0.0; remaining settings unknown | 66 terminal-work tasks in Terminal-Bench 4.0.0 | Published terminal-task resolution result | Exact harness version, fallback routing, complete budget and evaluation date unknown; descriptive result only | CO04;CO03 |
| Claude Fable 5.1 | Claude Code | Max effort; Harbor evaluation using Terminal-Bench 4.0.0; remaining settings unknown | 66 terminal-work tasks in Terminal-Bench 4.0.0 | Published terminal-task resolution result | Exact harness version, fallback routing, complete budget and evaluation date unknown; descriptive result only | CO05;CO03 |
| GLM-5.3 | Claude Code | Max effort; Harbor evaluation using Terminal-Bench 4.0.0; remaining settings unknown | 66 terminal-work tasks in Terminal-Bench 4.0.0 | Published terminal-task resolution result | Exact harness version, fallback routing, complete budget and evaluation date unknown; descriptive result only | CO06;CO03 |
| Claude Fable 5.1 | Claude Code | Endor Labs Agent Security League evaluation; exact inference settings unknown | Python tasks derived from CVE patches | Functional and security-test outcomes reported separately | Exact September run settings, adjusted denominator and uncertainty unknown | CO07 |
| gpt-6-astra | GitHub Copilot app/CLI | Business/Enterprise policy; exclusions and user budgets; human review | Existing GitHub team with approved hosted-data route | Central policy and credit controls; documented Astra route | Provider ZDR does not establish GitHub service retention; task advantage unmeasured | CH02;CH04;CH07;CH12;CH21 |
Measurements and prices
Benchmarks & cost
Benchmark evidence
Comparison depends on the dataset version, harness, evaluation budget, attempts, and metric. Vendor reports and independent measurements retain their source labels.
API price · output $/MTok
Output token list prices, not cost per accepted task. Input, caching, access route, and workload affect the total — see the data.
Who trains on your code
The privacy matrix
Green = recorded no-training policy; amber = choice or uncertainty; rose = recorded training default. Read the exact tier, route, retention, and exceptions in the full data. No training does not mean no retention.
| Provider | Product / tier | Trains by default? | Opt-out | Route / retention | Caveats |
|---|---|---|---|---|---|
| Gemini Developer API Paid Services | no | Paid terms; unpaid regional exceptions | Google API billing-linked project · Limited-period abuse logs; duration unspecified; Search grounding30days | Unpaid outside EEA/CH/UK can train; feedback separate | |
| DeepSeek | Open Platform API downstream route | unknown | Personal-data opt-out documented; API-wide coverage unverified | DeepSeek API · API-wide duration unknown | No blanket API no-training or ZDR commitment established |
| OpenAI | Standard API | no | unknown | Direct OpenAI API · Abuse logs up to 30 days with legal/harm exceptions; Responses state 30 days by default | Approved ZDR/MAM differs; store=false is not a blanket no-retention switch. |
| OpenAI | ChatGPT Business | no | unknown | Business workspace · unknown | Business no-training verified; consumer policy and exact service-retention duration not established here. |
| Anthropic | Claude Code commercial Team/Enterprise/API | no | unknown | Direct Anthropic commercial account · Standard 30 days; separately approved Enterprise ZDR; local transcripts normally 30 days | Fable-specific retention overrides generic expectations; feedback transcripts five years. |
| Anthropic | Claude consumer Free/Pro/Max | choice | yes | Direct consumer account · Training on: five years; off: 30 days | Account privacy preference applies to Claude Code; feedback retention separate. |
| Cursor | Privacy Mode; standard ZDR model | no | unknown | Cursor and selected provider · Provider ZDR with abuse exceptions; temporary encrypted file cache | Cloud Agents and designated non-ZDR models need separate assessment. |
| Cursor | Privacy Mode with Fable opt-in | no | unknown | Cursor to Anthropic · Provider retains prompts/outputs; exact duration unknown here | Team-wide approval required; Privacy Mode does not mean Fable ZDR. |
| GitHub | Copilot Business/Enterprise | no | unknown | GitHub; OpenAI/GitHub Azure · OpenAI provider ZDR agreement; GitHub service retention unknown | Provider agreement is not an end-to-end storage guarantee. |
| GitHub | Copilot Business/Enterprise Fable enabled | no | unknown | GitHub to Claude hosting providers · Anthropic retention by default; eligible enterprise temporary ZDR through 2026 | Off by default; exemptions require approval/internal-operations restrictions. |
| Z.ai | Enterprise/developer API Services under Additional Terms and DPA | no | Scoped content-use restriction applies unless explicitly agreed otherwise; no account control verified | Z.ai hosted API; applicable enterprise/developer agreement required · DPA says submitted/generated content is not stored on servers; other Customer Data retained for service or legal purposes, with no fixed duration established | Contractual commitments for the defined API scope, not an operational audit. Individual Coding Plan coverage unresolved. Human-access controls and metadata retention duration remain unknown. |
| Z.ai | Individual GLM Coding Plan | unknown | unknown | Z.ai hosted Coding Plan endpoint · unknown for this route | API terms provide content-use and storage protections for their defined scope. Their application to individual Coding Plan traffic is unverified; obtain account and endpoint-specific confirmation before using restricted repositories. |