Recommendation for Autocomplete

Code Completion

Our top recommendation for Code Completion, based on the public evidence we track, is Anthropic: Claude Opus 5.5.[1][2][3] Use when you need the highest-quality inline suggestions where benchmark-proven output quality outweighs cost concerns, as it tops LMArena WebDev (Elo 1818) and LiveBench Coding (89.25%). OpenAI: GPT-5.6 Sol is the next-ranked alternative. Use as a fallback when available, since LiveBench shows decent objective performance for code generation tasks.

About this recommendation

Updated
Sep 25, 2026
Evidence through
Sep 25, 2026
Sources
16
Revision
v85

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

94%

intended feed weight

Largest provider share

3 of 7

OpenAI

Provisional source breadth. 7 citation families and 0 practitioner families support the top result; 0 cautionary threads is retained. The largest citation family contributes 28%.

Sources evaluated

The task sets these weights before any model is scored.

winner: Claude Opus 5.5
Evaluation feedWeightWinner resultField measured
LiveBench Coding
30%
#117/20
LMArena WebDev
25%
#117/20
price weight
20%
97/10020/20
Aider Polyglot
15%
not measured3/20
OpenRouter usage
10%
87/10020/20

Provider concentration

Each exact model is scored separately; provider identity is not a ranking input.

OpenAI43%
  • OpenAI3 models
  • Anthropic2 models
  • 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.

RankModelRelative scoreCoveragePractitioner evidenceStrongest measured reason
01Claude Opus 5.5Anthropic
87
94%no linked practitioner threads#1 LiveBench Coding · #1 LMArena WebDev
02GPT-5.6 SolOpenAI
77
94%8 threads · 4 families · 4 cautions#4 LiveBench Coding · #15 LMArena WebDev
03GPT-5.6 LunaOpenAI
75
94%8 threads · 7 families · 2 cautions#6 LiveBench Coding · #31 LMArena WebDev
04GPT-6 SolOpenAI
74
94%4 threads · 3 families · 3 cautions#5 LMArena WebDev · #9 LiveBench Coding
05Claude Fable 5Anthropic
71
94%2 threads · 1 families · 1 cautions#3 LiveBench Coding · #11 LMArena WebDev
06GLM 5.2Z.ai
69
94%2 threads · 2 families · 0 cautions#16 LiveBench Coding · #17 LMArena WebDev
07Grok 4.7xAI
69
94%2 threads · 2 families · 0 cautions#9 LMArena WebDev · #30 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.

  1. Leads both major coding benchmarks, ranking #1 on LMArena WebDev and LiveBench Coding, indicating strong preference for its code completion quality in blind human evaluations.

    Best when: Use when you need the highest-quality inline suggestions where benchmark-proven output quality outweighs cost concerns, as it tops LMArena WebDev (Elo 1818) and LiveBench Coding (89.25%).

    Tips

    • Use when you need the highest-quality inline suggestions where benchmark-proven output quality outweighs cost concerns, as it tops LMArena WebDev (Elo 1818) and LiveBench Coding (89.25%).
      Source 1
      “Ranks #1 of 91 on LMArena's WebDev coding arena (Elo 1818), a leaderboard built from blind human preference votes on coding tasks.”
      LMArena WebDev (coding) arenaOpen original ↗
      Source 2
      “Scores 89.25% on LiveBench Coding (#1 of 58), an objective evaluation of code generation and completion.”
      LiveBench CodingOpen original ↗
  2. Scores 83.94% on LiveBench Coding (#4) but suffers from repeated capacity failures and user reports of quality regression in newer versions.

    Best when: Use as a fallback when available, since LiveBench shows decent objective performance for code generation tasks.

    Tips

    • Use as a fallback when available, since LiveBench shows decent objective performance for code generation tasks.
      Source 3
      “Scores 83.94% on LiveBench Coding (#4 of 58), an objective evaluation of code generation and completion.”
      LiveBench CodingOpen original ↗

    Watch out for

    • Expect intermittent 503 capacity errors and 'Selected model is at capacity' failures that block completion, reported across multiple platforms and dates.
      Source 4
      “### What version of Codex CLI is running? 0.154.0 ### What subscription do you have? ChatGPT Pro 20x ### Which model were you using? gpt-6-astra, gpt-5.6-sol, gpt-5.6-luna, gpt-5.6-terra, gpt-5.5 ### What platform is your computer? Darwin 25.6.0 arm64 arm ### What terminal emulator and version are you using (if applicable)? zsh ### Codex doctor report ### What issue are you seeing? Repeated ⚠ Selected model is at capacity. Please try a different model. Trying different model did not help. ### W…”
      Source 5
      “I’m seeing the same capacity failure today (Sep 10, 2026) on **GPT-5.6 Sol** on Windows. What I observed: - A trivial prompt such as `test` / `tes` repeatedly fails with: `Selected model is at capacity. Please try a different model.` - In Codex Desktop through my local Responses-compatible router, the upstream failure is surfaced as: `503 Service Unavailable` / `Our servers are currently overloaded. Please try again later.` - I verified the local router itself is reachable (`127.0.0.1:20128`, T…”
    • Note user reports that GPT-6 Sol (marketed as successor) produces 'epic levels of slop,' with experienced users reverting to 5.6 Sol for quality.
      Source 6
      “Update: Sol 6 is a heaping pile of garbage. Just epic levels of slop. And r codex etc is full of people noticing the same. I've switched back to 5.6 Sol. What they're selling as Sol 6 is really what would have been Terra before, and it's awful.”
      cmrdporcupineOpen original ↗
  3. Ranks #6 on LiveBench Coding (82.92%) but exhibits API compatibility issues and subscription gating mismatches that complicate deployment.

    Best when: Consider for code completion where LiveBench scores indicate acceptable generation quality for mid-tier performance needs.

    Tips

    • Consider for code completion where LiveBench scores indicate acceptable generation quality for mid-tier performance needs.
      Source 7
      “Scores 82.92% on LiveBench Coding (#6 of 58), an objective evaluation of code generation and completion.”
      LiveBench CodingOpen original ↗

    Watch out for

    • Handle API integration carefully, as the upstream rejects any `max_tokens` field presence on `/v1/responses` (even null), requiring field omission rather than standard parameter handling.
      Source 8
      “## 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…”
    • Verify subscription tiers manually, as the model is marked `paid` in gate tables but was observed serving Free accounts, causing priority ranking mismatches.
      Source 9
      “### Description omp's static `openai-codex` plan-requirement table does not match what the Codex backend accepted from ChatGPT **Free** accounts. I tested every `openai-codex` model omp ships, plus `gpt-5.6` from the gate table, on two Free and two Team accounts on 2026-09-25. There are two kinds of mismatch: 1. **Gated but served:** `gpt-5.6-luna` is marked `paid`, yet both Free accounts listed it and completed turns with it. omp ranks Free accounts last for this model (`planPriority = 2`) eve…”
  4. Ranks #5 on LMArena WebDev (Elo 1686) but faces severe user quality complaints and version confusion, with reports that GPT-6 Sol is rebranded inferior material.

    Best when: Consider where LMArena ranking suggests reasonable preference performance in blind evaluations.

    Tips

    • Consider where LMArena ranking suggests reasonable preference performance in blind evaluations.
      Source 10
      “Ranks #5 of 91 on LMArena's WebDev coding arena (Elo 1686), a leaderboard built from blind human preference votes on coding tasks.”
      LMArena WebDev (coding) arenaOpen original ↗

    Watch out for

    • Heed widespread user reports that 'Sol 6 is a heaping pile of garbage' and 'epic levels of slop,' with the community noting it appears to be rebranded Terra-grade material rather than a true successor.
      Source 6
      “Update: Sol 6 is a heaping pile of garbage. Just epic levels of slop. And r codex etc is full of people noticing the same. I've switched back to 5.6 Sol. What they're selling as Sol 6 is really what would have been Terra before, and it's awful.”
      cmrdporcupineOpen original ↗
  5. Ranks third on LiveBench Coding (85.99%) and maintains competitive placement on LMArena WebDev (#11), offering a middle ground between top-tier quality and practical deployment.

    Best when: Consider for cost-conscious workflows where LiveBench shows strong objective performance (#3 of 58) without the top-tier pricing of benchmark leaders.

    Tips

    • Consider for cost-conscious workflows where LiveBench shows strong objective performance (#3 of 58) without the top-tier pricing of benchmark leaders.
      Source 11
      “Scores 85.99% on LiveBench Coding (#3 of 58), an objective evaluation of code generation and completion.”
      LiveBench CodingOpen original ↗

    Watch out for

    • Watch for model self-identification errors where the system incorrectly denies its own model name exists, which could affect logging or debugging workflows.
      Source 12
      “+1. And there had been a long standing description for a product like this: not fit for purpose. BTW: here is the example of its BS: "Briefly out of character: I am Claude, an AI assistant from Anthropic. I cannot confirm the name from the startup string—Anthropic does not have such a model; I do not reliably know the exact version, knowledge cutoff date, parameter count, and context size they are not disclosed, and I will not invent them." This "Anthropic does not have such a model" seems to m…”
  6. Open-weight option ranking #17 on LMArena WebDev with reported cost efficiency for pair-programming workflows, though benchmark scores trail closed alternatives.

    Best when: Deploy on budget-conscious inference providers like Cerebras for pair-programming workflows where you maintain active oversight, as users report delivering production work at low token costs.

    Tips

    • Deploy on budget-conscious inference providers like Cerebras for pair-programming workflows where you maintain active oversight, as users report delivering production work at low token costs.
      Source 13
      “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.”

    Watch out for

    • Accept lower benchmark standing (#17 on LMArena WebDev, Elo 1600) compared to closed-weight leaders, which may impact suggestion quality on complex completions.
      Source 14
      “Ranks #17 of 91 on LMArena's WebDev coding arena (Elo 1600), a leaderboard built from blind human preference votes on coding tasks.”
      LMArena WebDev (coding) arenaOpen original ↗
  7. Places #9 on LMArena WebDev (Elo 1632) with reported speed advantages in practice, though tooling integration gaps limit controllability.

    Best when: Use for rapid iteration cycles where the 'fast option' allows quick correction loops, particularly in frontend workflows where users report it outpaces alternatives.

    Tips

    • Use for rapid iteration cycles where the 'fast option' allows quick correction loops, particularly in frontend workflows where users report it outpaces alternatives.
      Source 15
      “I'm curious: what languages or frameworks is this in? The Django code that comes out of composer2.5, to me, was insulting. Grok definitely was a step up, especially because the fast option reaaally is fast so even if it came out a bit wrong I could just whip it into perfection. For frontend work, it's a different story. You can still tell that composer2.5 is taking the long route, but I don't think it's as egregious as with Django. Also, composer2.5 would routinely run commands that were really…”

    Watch out for

    • Avoid if you need fine-grained thinking control via SDK, as the pi-cursor-sdk lacks mapping for Grok 4.7's thinking levels despite Cursor supporting them natively.
      Source 16
      “## Summary On published `pi-cursor-sdk@0.3.9`, **Grok 4.7 is selectable** in `/model` and `pi --list-models`, but **pi-controllable thinking is unavailable** (`thinking=no`). Cursor already exposes a thinking control for this model; the extension does not map it. This is an extension mapping gap, not a Cursor “Grok 4.7 has no reasoning” bug. The model can still think on the Cursor side. Pi cannot drive the level with `shift+tab`, `--thinking`, or a final `:low|:medium|:high|:xhigh` suffix. ## E…”

Frequently asked

What is the top-ranked model for Code Completion?
Anthropic: Claude Opus 5.5 ranks first in the current evidence-weighted comparison. Use when you need the highest-quality inline suggestions where benchmark-proven output quality outweighs cost concerns, as it tops LMArena WebDev (Elo 1818) and LiveBench Coding (89.25%).[1][2]
What is an alternative to Anthropic: Claude Opus 5.5?
OpenAI: GPT-5.6 Sol is the next-ranked option. Use as a fallback when available, since LiveBench shows decent objective performance for code generation tasks.[3]

Sources

  1. 1

    “Ranks #1 of 91 on LMArena's WebDev coding arena (Elo 1818), a leaderboard built from blind human preference votes on coding tasks.”

    LMArena WebDev (coding) arena · Benchmark · Sep 23, 2026
  2. 2

    “Scores 89.25% on LiveBench Coding (#1 of 58), an objective evaluation of code generation and completion.”

    LiveBench Coding · Benchmark · Jun 25, 2026
  3. 3

    “Scores 83.94% on LiveBench Coding (#4 of 58), an objective evaluation of code generation and completion.”

    LiveBench Coding · Benchmark · Jun 25, 2026
  4. 4

    “### What version of Codex CLI is running? 0.154.0 ### What subscription do you have? ChatGPT Pro 20x ### Which model were you using? gpt-6-astra, gpt-5.6-sol, gpt-5.6-luna, gpt-5.6-terra, gpt-5.5 ### What platform is your computer? Darwin 25.6.0 arm64 arm ### What terminal emulator and version are you using (if applicable)? zsh ### Codex doctor report ### What issue are you seeing? Repeated ⚠ Selected model is at capacity. Please try a different model. Trying different model did not help. ### W…”

    sunzh233 · GitHub · Sep 10, 2026
  5. 5

    “I’m seeing the same capacity failure today (Sep 10, 2026) on **GPT-5.6 Sol** on Windows. What I observed: - A trivial prompt such as `test` / `tes` repeatedly fails with: `Selected model is at capacity. Please try a different model.` - In Codex Desktop through my local Responses-compatible router, the upstream failure is surfaced as: `503 Service Unavailable` / `Our servers are currently overloaded. Please try again later.` - I verified the local router itself is reachable (`127.0.0.1:20128`, T…”

    FawziLinggo · GitHub · Sep 10, 2026
  6. 6

    “Update: Sol 6 is a heaping pile of garbage. Just epic levels of slop. And r codex etc is full of people noticing the same. I've switched back to 5.6 Sol. What they're selling as Sol 6 is really what would have been Terra before, and it's awful.”

    cmrdporcupine · Hacker News · Sep 23, 2026
  7. 7

    “Scores 82.92% on LiveBench Coding (#6 of 58), an objective evaluation of code generation and completion.”

    LiveBench Coding · Benchmark · Jun 25, 2026
  8. 8

    “## 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
  9. 9

    “### Description omp's static `openai-codex` plan-requirement table does not match what the Codex backend accepted from ChatGPT **Free** accounts. I tested every `openai-codex` model omp ships, plus `gpt-5.6` from the gate table, on two Free and two Team accounts on 2026-09-25. There are two kinds of mismatch: 1. **Gated but served:** `gpt-5.6-luna` is marked `paid`, yet both Free accounts listed it and completed turns with it. omp ranks Free accounts last for this model (`planPriority = 2`) eve…”

    aydgn · GitHub · Sep 25, 2026
  10. 10

    “Ranks #5 of 91 on LMArena's WebDev coding arena (Elo 1686), a leaderboard built from blind human preference votes on coding tasks.”

    LMArena WebDev (coding) arena · Benchmark · Sep 23, 2026
  11. 11

    “Scores 85.99% on LiveBench Coding (#3 of 58), an objective evaluation of code generation and completion.”

    LiveBench Coding · Benchmark · Jun 25, 2026
  12. 12

    “+1. And there had been a long standing description for a product like this: not fit for purpose. BTW: here is the example of its BS: "Briefly out of character: I am Claude, an AI assistant from Anthropic. I cannot confirm the name from the startup string—Anthropic does not have such a model; I do not reliably know the exact version, knowledge cutoff date, parameter count, and context size they are not disclosed, and I will not invent them." This "Anthropic does not have such a model" seems to m…”

    pbgcp2026 · Hacker News · Jun 12, 2026
  13. 13

    “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
  14. 14

    “Ranks #17 of 91 on LMArena's WebDev coding arena (Elo 1600), a leaderboard built from blind human preference votes on coding tasks.”

    LMArena WebDev (coding) arena · Benchmark · Sep 23, 2026
  15. 15

    “I'm curious: what languages or frameworks is this in? The Django code that comes out of composer2.5, to me, was insulting. Grok definitely was a step up, especially because the fast option reaaally is fast so even if it came out a bit wrong I could just whip it into perfection. For frontend work, it's a different story. You can still tell that composer2.5 is taking the long route, but I don't think it's as egregious as with Django. Also, composer2.5 would routinely run commands that were really…”

    jorl17 · Hacker News · Sep 22, 2026
  16. 16

    “## Summary On published `pi-cursor-sdk@0.3.9`, **Grok 4.7 is selectable** in `/model` and `pi --list-models`, but **pi-controllable thinking is unavailable** (`thinking=no`). Cursor already exposes a thinking control for this model; the extension does not map it. This is an extension mapping gap, not a Cursor “Grok 4.7 has no reasoning” bug. The model can still think on the Cursor side. Pi cannot drive the level with `shift+tab`, `--thinking`, or a final `:low|:medium|:high|:xhigh` suffix. ## E…”

    joaomelga · GitHub · Sep 22, 2026

Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.