Recommendation for Autocomplete
Code Completion
Our top recommendation for Code Completion, based on the public evidence we track, is OpenAI: GPT-5.6 Luna.[1][2] Watch out: Strip any max_tokens field from API requests entirely, as this model rejects even null values with HTTP 400 errors on /v1/responses endpoints. OpenAI: GPT-5.6 Sol is the next-ranked alternative. Its currently supported evidence is cautionary: Expect occasional long recovery pauses during sessions, with reports of 10+ minute stalls requiring reconnection in Codex CLI usage.
About this recommendation
- Updated
- Sep 6, 2026
- Evidence through
- Sep 6, 2026
- Sources
- 11
- Revision
- v66
Decision audit
Why this result
Inspect the inputs and the computed order behind the recommendation.
Models screened
20
live candidates
Evaluation feeds
5
task-weighted
Winner coverage
93%
intended feed weight
Largest provider share
3 of 8
Anthropic
Sources evaluated
The task sets these weights before any model is scored.
| Evaluation feed | Weight | Winner result | Field measured |
|---|---|---|---|
| LiveBench Coding | 30% | #5 | 16/20 |
| LMArena WebDev | 25% | #29 | 17/20 |
| price weight | 20% | 100/100 | 20/20 |
| Aider Polyglot | 15% | not measured | 3/20 |
| OpenRouter usage | 10% | 99/100 | 20/20 |
Provider concentration
Each exact model is scored separately; provider identity is not a ranking input.
- Anthropic3 models
- OpenAI2 models
- tencent1 model
- xAI1 model
- Z.ai1 model
Decision table
Every published model is shown in computed order. Practitioner sources are distinct community threads, not the citations repeated in the prose below.
| Rank | Model | Relative score | Coverage | Practitioner evidence | Strongest measured reason |
|---|---|---|---|---|---|
| 01 | GPT-5.6 LunaOpenAI | 78 | 93% | 5 threads · 5 families · 1 cautions | #5 LiveBench Coding · #29 LMArena WebDev |
| 02 | GPT-5.6 SolOpenAI | 75 | 93% | 2 threads · 2 families · 1 cautions | #3 LiveBench Coding · #11 LMArena WebDev |
| 03 | Claude Fable 5Anthropic | 75 | 93% | 2 threads · 1 families · 1 cautions | #2 LiveBench Coding · #7 LMArena WebDev |
| 04 | GLM 5.2Z.ai | 72 | 93% | 2 threads · 2 families · 0 cautions | #13 LiveBench Coding · #15 LMArena WebDev |
| 05 | Claude Opus 4.8Anthropic | 69 | 93% | 1 threads · 1 families · 1 cautions | #7 LiveBench Coding · #19 LMArena WebDev |
| 06 | Claude Sonnet 5Anthropic | 68 | 93% | 2 threads · 2 families · 1 cautions | #11 LiveBench Coding · #25 LMArena WebDev |
| 07 | Hy4 previewtencent | 68 | 72% | no linked practitioner threads | #10 LMArena WebDev |
| 08 | Grok 4.6xAI | 68 | 93% | 1 threads · 1 families · 0 cautions | #8 LMArena WebDev · #28 LiveBench Coding |
Relative score combines normalized benchmark quality and signal coverage; independent practitioner evidence and freshness are bounded tie-breakers. It is an ordering score, not an absolute quality percentage. The writing model receives this order and cannot change it.
GPT-5.6 Luna scores 82.92% on LiveBench Coding and ranks twenty-ninth on LMArena WebDev, showing solid objective performance but weaker human preference rankings than its Sol variant.
Best when: Consider only after reviewing the cited caution.
Watch out for
- Strip any `max_tokens` field from API requests entirely, as this model rejects even null values with HTTP 400 errors on `/v1/responses` endpoints.
- Prepare for request blocking that halts workflow, with reports of unexplained blocks that prevent meaningful work resumption.
GPT-5.6 Sol ranks third on LiveBench Coding at 83.94% and eleventh on LMArena WebDev, offering strong objective performance with broad deployment availability through Bedrock, Azure, and OpenRouter.
Best when: Consider only after reviewing the cited caution.
Watch out for
- Expect occasional long recovery pauses during sessions, with reports of 10+ minute stalls requiring reconnection in Codex CLI usage.
Anthropic: Claude Fable 5 scores 85.99% on LiveBench Coding (#2 of 52), an objective evaluation of code generation and completion.
Best when: Consider only after reviewing the cited caution.
GLM 5.2 scores 79.65% on LiveBench Coding and ranks sixteenth on LMArena WebDev, with community reports of practical utility for pair programming at low cost when used with fast inference providers.
Best when: Deploy with Cerebras or other fast inference providers for low-cost pair programming assistance, as engineers report delivering production work with GLM 4.7 and similar variants at favorable pricing.
Tips
- Deploy with Cerebras or other fast inference providers for low-cost pair programming assistance, as engineers report delivering production work with GLM 4.7 and similar variants at favorable pricing.
Watch out for
- Expect lower absolute benchmark performance at 79.65% LiveBench Coding, placing fourteenth of fifty-two, below all major closed-weight alternatives.
Claude Opus 4.8 scores 81.83% on LiveBench Coding and ranks nineteenth on LMArena WebDev, showing decent objective performance but weaker human preference than the Fable and Sonnet variants.
Best when: Use when you need Anthropic's largest model variant with established benchmark presence at 81.83% LiveBench accuracy.
Tips
- Use when you need Anthropic's largest model variant with established benchmark presence at 81.83% LiveBench accuracy.
Watch out for
- Confirm availability before integration, as CLI probes in late July 2026 returned 'not available' errors for this model ID despite appearing in official model lists.
Claude Sonnet 5 appears in CLI model lists but lacks direct benchmark scores in the evidence, with availability confirmed across Anthropic's standard deployment routes.
Best when: Consider only after reviewing the cited caution.
Watch out for
- Verify actual availability before committing, as CLI probes returned 'not available' errors for this model ID in late July 2026 testing despite appearing in model lists.
Hy4 preview ranks tenth on LMArena WebDev with an Elo of 1621, showing competitive human preference performance as an open-weight option from Tencent.
Best when: Evaluate as an open-weight alternative with strong human preference signals, ranking tenth on LMArena WebDev coding arena.
Tips
- Evaluate as an open-weight alternative with strong human preference signals, ranking tenth on LMArena WebDev coding arena.
Grok 4.6 ranks eighth on LMArena WebDev with strong human preference signals and is noted as cheap on Cursor Pro, though its LiveBench Coding score of 76.78% lags behind top performers.
Best when: Prioritize for budget-conscious deployments where cost trumps peak accuracy, as it is reported as 'good enough and cheap' on the $20 Cursor Pro plan.
Tips
- Prioritize for budget-conscious deployments where cost trumps peak accuracy, as it is reported as 'good enough and cheap' on the $20 Cursor Pro plan.
- Use when human preference rankings matter, placing eighth on LMArena WebDev with an Elo of 1625.
Watch out for
- Accept lower objective code quality, scoring 76.78% on LiveBench Coding and ranking thirtieth of fifty-two models evaluated.
Frequently asked
- What should I watch out for with OpenAI: GPT-5.6 Luna?
- Strip any `max_tokens` field from API requests entirely, as this model rejects even null values with HTTP 400 errors on `/v1/responses` endpoints.[1]
Sources
- 1
“## Summary PR #330 made per-provider `max_tokens` caps land on every path. It could not help `gpt-5.6-luna` because that upstream rejects the *presence* of the `max_tokens` field on `/v1/responses` — even `null` returns HTTP 400. The only fix is to omit the field entirely. This branch adds a declarative opt-in flag (TOML field + env var) that lets the user mark specific providers as "drop the `max_tokens` field from the wire body", and threads it through `OpenCodeGoResponsesProvider`. ## What c…”
airvzxf · GitHub · Aug 10, 2026 - 2
“### What version of Codex CLI is running? codex-cli 0.146.0 ### What subscription do you have? Plus ### Which model were you using? gpt-5.6-Sol ### What platform is your computer? Linux 7.0.0-28-generic x86_64 x86_64 ### What terminal emulator and version are you using (if applicable)? just terminal ### Codex doctor report ### What issue are you seeing? Multiple times over the last few sessions since updating codex to the recent version, I have had codex stop, and spend 10 minutes or more re-co…”
Hermetica · GitHub · Aug 2, 2026 - 3
“### What version of Codex CLI is running? codex-cli 0.145.0 ### What subscription do you have? Pro ### Which model were you using? gpt-5.6-luna ### What platform is your computer? Ubuntu ### What terminal emulator and version are you using (if applicable)? _No response_ ### Codex doctor report ### What issue are you seeing? request blocked for unknown reason and cannot resume work in any meaningful way. ### What steps can reproduce the bug? Uploaded thread: 019f3652-8da7-79d3-973b-4397356f6f36…”
Athena-Noesis · GitHub · Jul 28, 2026 - 4
“If you are a real engineer and uses the LLM as a pair programmer instead of delegating everything to it, even GLM 4.7 was already good enough to help you with a lot of work. I used it with Cerebras inference at a time when it had a good coding plan at a low price, and delivered tons of stuff using it.”
elzbardico · Hacker News · Jun 22, 2026 - 5
“Scores 79.65% on LiveBench Coding (#14 of 52), an objective evaluation of code generation and completion.”
LiveBench Coding · Benchmark · Jun 25, 2026 - 6
“Scores 81.83% on LiveBench Coding (#8 of 52), an objective evaluation of code generation and completion.”
LiveBench Coding · Benchmark · Jun 25, 2026 - 7
“## What was measured, 26-Jul-2026 Asked `copilot -p` to list the models available via `--model`. It answered: `grok-4.5`, `claude-opus-4-8`, `claude-opus-4-7`, `claude-sonnet-5`, `claude-sonnet-4-6`, `claude-haiku-4-5`, plus aliases, and `DEFAULT=gpt-5-mini`. Then probed each id directly. **Every one returned `Model " " from --model flag is not available`** — including `gpt-5-mini` and `grok-4.5`, the two the CLI had just named. Probed: gpt-5-mini, gpt-5.1, gpt-5.1-codex, gpt-4.1, o4-mini, gemi…”
ajmccaus · GitHub · Jul 26, 2026 - 8
“Ranks #10 of 89 on LMArena's WebDev coding arena (Elo 1621), a leaderboard built from blind human preference votes on coding tasks.”
LMArena WebDev (coding) arena · Benchmark · Sep 5, 2026 - 9
“$20 cursor pro plan only, grok 4.6 is good enough and cheap”
armadillo1354 · Hacker News · Aug 14, 2026 - 10
“Ranks #8 of 89 on LMArena's WebDev coding arena (Elo 1625), a leaderboard built from blind human preference votes on coding tasks.”
LMArena WebDev (coding) arena · Benchmark · Sep 5, 2026 - 11
“Scores 76.78% on LiveBench Coding (#30 of 52), an objective evaluation of code generation and completion.”
LiveBench Coding · Benchmark · Jun 25, 2026
Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.