Recommendation for Cheapest capable

Cheapest Capable LLM

Our top recommendation for Cheapest Capable LLM, based on the public evidence we track, is Z.ai: GLM 5.2.[1][2] Route through Vercel AI Gateway or similar aggregators to access sub-dollar per-million-token pricing on an open-weight model, bypassing the $1.40/$4.40 direct Z.ai rates. Watch out: Expect zero-cost reporting in some cost-tracking tools; GLM 5.2 rows are absent from default pricing catalogs, causing spend calculations to silently drop to $0 regardless of actual usage. Google: Gemini 3.5 Flash Lite is the next-ranked alternative. Use as a quota-fallback in free-tier applications; it automatically substitutes when Gemini 3.6 Flash hits its 20-requests-per-day project limit, keeping onboarding flows alive.

About this recommendation

Updated
Sep 4, 2026
Evidence through
Sep 4, 2026
Sources
9
Revision
v56

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

67%

intended feed weight

Largest provider share

3 of 6

Anthropic

Provisional source breadth. 10 citation families and 3 practitioner families support the top result; 0 cautionary threads is retained. The largest citation family contributes 23%.

Sources evaluated

The task sets these weights before any model is scored.

winner: GLM 5.2
Evaluation feedWeightWinner resultField measured
price weight
40%
99/10020/20
Route reliabilityunavailable
30%
feed unavailable0/20
LiveBench Instruction Following
15%
#4017/20
OpenRouter usage
10%
97/10020/20
LMArena Text
5%
#2017/20

Provider concentration

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

Anthropic50%
  • Anthropic3 models
  • Google1 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.

RankModelRelative scoreCoveragePractitioner evidenceStrongest measured reason
01GLM 5.2Z.ai
71
67%3 threads · 3 families · 0 cautions#20 LMArena Text · #40 LiveBench Instruction Following
02Gemini 3.5 Flash LiteGoogle
71
67%5 threads · 4 families · 2 cautions#24 LiveBench Instruction Following · #32 LMArena Text
03Claude Haiku 4.5Anthropic
70
61%4 threads · 4 families · 1 cautions#71 LMArena Text
04Grok 4.6xAI
70
67%4 threads · 2 families · 2 cautions#15 LiveBench Instruction Following · #27 LMArena Text
05Claude Fable 5Anthropic
68
67%1 threads · 1 families · 0 cautions#1 LMArena Text · #5 LiveBench Instruction Following
06Claude Opus 4.6Anthropic
66
67%no linked practitioner threads#2 LMArena Text · #34 LiveBench Instruction Following

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. GLM 5.2 is an open-weight model with gateway-dependent pricing that can run as low as $0.80/$2.52 per 1M tokens through Vercel AI Gateway, though catalog coverage in cost-tracking tools remains inconsistent.

    Best when: Route through Vercel AI Gateway or similar aggregators to access sub-dollar per-million-token pricing on an open-weight model, bypassing the $1.40/$4.40 direct Z.ai rates.

    Tips

    • Route through Vercel AI Gateway or similar aggregators to access sub-dollar per-million-token pricing on an open-weight model, bypassing the $1.40/$4.40 direct Z.ai rates.
      Source 1
      Verified on 0.50.0: these real, executable models price at $0 (`unknown model → cost set to $0`) regardless of spelling — the rows are simply absent: - `zai/glm-5.2` — $1.40 / $4.40 per MTok in/out ([Z.ai pricing](https://docs.z.ai/guides/overview/pricing); gateway-routed rates run lower, e.g. $0.80/$2.52 via [Vercel AI Gateway](https://vercel.com/ai-gateway/models/glm-5.2)) - `xai/grok-4.5` — $2 / $6 per MTok ([xAI models & pricing](https://docs.x.ai/docs/models); base tier for <200k-token pro…

    Watch out for

    • Expect zero-cost reporting in some cost-tracking tools; GLM 5.2 rows are absent from default pricing catalogs, causing spend calculations to silently drop to $0 regardless of actual usage.
      Source 1
      Verified on 0.50.0: these real, executable models price at $0 (`unknown model → cost set to $0`) regardless of spelling — the rows are simply absent: - `zai/glm-5.2` — $1.40 / $4.40 per MTok in/out ([Z.ai pricing](https://docs.z.ai/guides/overview/pricing); gateway-routed rates run lower, e.g. $0.80/$2.52 via [Vercel AI Gateway](https://vercel.com/ai-gateway/models/glm-5.2)) - `xai/grok-4.5` — $2 / $6 per MTok ([xAI models & pricing](https://docs.x.ai/docs/models); base tier for <200k-token pro…
  2. Gemini 3.5 Flash Lite offers the lowest per-token pricing in Google's current catalog at $0.30/$2.50 per 1M, serving as a reliable fallback when quota limits block higher-tier Flash models.

    Best when: Use as a quota-fallback in free-tier applications; it automatically substitutes when Gemini 3.6 Flash hits its 20-requests-per-day project limit, keeping onboarding flows alive.

    Tips

    • Use as a quota-fallback in free-tier applications; it automatically substitutes when Gemini 3.6 Flash hits its 20-requests-per-day project limit, keeping onboarding flows alive.
      Source 2
      ## Who As a PairPocket user on the Gemini free tier, I want the app to prefer `gemini-3.6-flash` and automatically use `gemini-3.5-flash-lite` while 3.6 is quota-blocked, so that onboarding and receipt scans keep working without waiting on a failed 3.6 call every time. ## Why Free-tier limits are per project and per model row (RPM, TPM, RPD). RPD resets at midnight Pacific Time. RPM/TPM use a rolling window. Hard-coding quota numbers is unreliable; AI Studio shows live limits. Trying 3.6 on eve…
      minsikpaul92Open original ↗

    Watch out for

    • Avoid setting thinking_budget=0 in your API calls; Flash Lite rejects this configuration with 400 INVALID_ARGUMENT, unlike other Gemini models that accept it.
      Source 3
      Implementation note — the premise of this issue was wrong in one respect, and it would have taken prod down. `gemini-3.5-flash-lite` **rejects `thinking_budget=0`** with `400 INVALID_ARGUMENT`. `GeminiProvider._build_config` (`backend/services/llm.py`) sets that unconditionally, so a pure env-var switch would have failed *every* Gemini call rather than degrading quality. Verified against the live API on 2026-08-02: | config | result on `gemini-3.5-flash-lite` | |---|---| | `thinkingBudget: 0` |…
  3. Claude Haiku 4.5 provides validated JSON structured output with competitive latency, though cost-routing misconfigurations in some systems have historically billed it at Opus rates.

    Best when: Use for assistant extraction calls requiring structured JSON output; it offers reliable schema adherence with lower per-turn cost than Sonnet or Opus tiers for typical 1.5K input / 300 output token workloads.

    Tips

    • Use for assistant extraction calls requiring structured JSON output; it offers reliable schema adherence with lower per-turn cost than Sonnet or Opus tiers for typical 1.5K input / 300 output token workloads.
      Source 4
      ## Question Which models are candidates for the Assistant's extraction call, and what does one Turn cost on each? Compare Claude Haiku 4.5 (and any cheaper current Claude tier) called directly against two or three cheap models reachable via OpenRouter. For each: price per million input and output tokens, cost of a typical Turn (about 1.5K input tokens of system prompt plus conversation, about 300 output tokens of structured JSON), structured-output or JSON-schema support, typical latency, and r…

    Watch out for

    • Audit billing records carefully; production incidents show Haiku 4.5 calls being silently upcharged to Opus 4.8 rates (5× cost) or misrouted to thinking variants due to alias resolution bugs in workflow compilers.
      Source 5
      ## Objective Find and fix why Avenger (`.github/workflows/avenger.md`) declares `model: claude-haiku-4.5` but audited runs execute on `claude-opus-4-8`. ## Issue context Issue #50312 found the largest cost driver on the Avenger workflow is a model mismatch: frontmatter/intended budget points to Haiku, but `agent_usage.json` shows Opus 4.8 in all 3 sampled runs (5x more expensive per token). This suggests a compiler/runtime bug in model propagation, alias resolution, or engine argument generatio…
      github-actions[bot]Open original ↗
      Source 6
      ## The incident that surfaced this While pinning a cheap model for WhatsApp channel turns (chat-surface cost reduction, 2026-08-06), we set `ExecutionConfig.model_name: "composer-2.5"` and verified the pin held — but production billing records show every call resolved to **`composer-2.5-fast`**, which carries ~6x the base input/output rates ($3.00/$15.00 vs $0.50/$2.50 per M tokens). The same happened to the other calibration arm: `claude-haiku-4-5` billed as `claude-4.5-haiku-thinking`. Root c…
  4. Grok 4.6 ranks #29 on LMArena with Elo 1461 and offers generous usage through Cursor subscription, though real-world benchmark costs run higher than per-token rates suggest due to elevated token consumption.

    Best when: Use via Cursor Pro subscription for effectively capped costs; the $20 plan includes generous Grok 4.6 access that can undercut metered API pricing for heavy interactive use.

    Tips

    • Use via Cursor Pro subscription for effectively capped costs; the $20 plan includes generous Grok 4.6 access that can undercut metered API pricing for heavy interactive use.
      Source 7
      $20 cursor pro plan only, grok 4.6 is good enough and cheap
      armadillo1354Open original ↗

    Watch out for

    • Do not assume per-token affordability translates to cheap real-world runs; Artificial Analysis measured $1,068 to complete their suite versus $485 for Gemini 3.7 Flash, indicating higher actual token consumption than competitors.
      Source 8
      Artificial Analysis shows Grok 4.6 taking $1,068 to run their suite while Gemini 3.7 Flash takes $485. So it looks like Gemini 3.7 Flash is less than half the price in the real world. Per-token cost isn't a great metric given that some use way more tokens than others.
  5. Claude Fable 5 ranks #1 on LMArena with Elo 1507, representing the quality frontier.

    Best when: Use for final review stages where maximum quality justifies cost; Forge routes only major-version Dependabot PRs to Fable 5, keeping minors on cheaper Sonnet 4.5 to control spend.

    Tips

    • Use for final review stages where maximum quality justifies cost; Forge routes only major-version Dependabot PRs to Fable 5, keeping minors on cheaper Sonnet 4.5 to control spend.
      Source 9
      ## What Forge now cost-routes **Dependabot patch/minor** PR reviews to a cheaper model (`claude-sonnet-4-5`), keeping majors and everything else on the fleet default (`claude-fable-5`). The routing decision is made **server-side** in Forge (`/api/v1/reviewer-models`, shipped in TrueTime-Health/forge#176) — but it can only route when the calling workflow hands Forge the PR's author and title. This repo's `claude-code-review.yml` doesn't yet, so its Dependabot PRs still bill the expensive reviewe…
      truetime-helper[bot]Open original ↗
  6. Anthropic: Claude Opus 4.6 ranks #2 of 144 on LMArena's overall text arena (Elo 1505), based on blind human preference votes.

    Best when: Consider only after reviewing the cited caution.

Frequently asked

What is the top-ranked model for Cheapest Capable LLM?
Z.ai: GLM 5.2 ranks first in the current evidence-weighted comparison. Route through Vercel AI Gateway or similar aggregators to access sub-dollar per-million-token pricing on an open-weight model, bypassing the $1.40/$4.40 direct Z.ai rates.[1]
What should I watch out for with Z.ai: GLM 5.2?
Expect zero-cost reporting in some cost-tracking tools; GLM 5.2 rows are absent from default pricing catalogs, causing spend calculations to silently drop to $0 regardless of actual usage.[1]
What is an alternative to Z.ai: GLM 5.2?
Google: Gemini 3.5 Flash Lite is the next-ranked option. Use as a quota-fallback in free-tier applications; it automatically substitutes when Gemini 3.6 Flash hits its 20-requests-per-day project limit, keeping onboarding flows alive.[2]

Sources

  1. 1

    Verified on 0.50.0: these real, executable models price at $0 (`unknown model → cost set to $0`) regardless of spelling — the rows are simply absent: - `zai/glm-5.2` — $1.40 / $4.40 per MTok in/out ([Z.ai pricing](https://docs.z.ai/guides/overview/pricing); gateway-routed rates run lower, e.g. $0.80/$2.52 via [Vercel AI Gateway](https://vercel.com/ai-gateway/models/glm-5.2)) - `xai/grok-4.5` — $2 / $6 per MTok ([xAI models & pricing](https://docs.x.ai/docs/models); base tier for <200k-token pro…

    lra · GitHub · Aug 1, 2026
  2. 2

    ## Who As a PairPocket user on the Gemini free tier, I want the app to prefer `gemini-3.6-flash` and automatically use `gemini-3.5-flash-lite` while 3.6 is quota-blocked, so that onboarding and receipt scans keep working without waiting on a failed 3.6 call every time. ## Why Free-tier limits are per project and per model row (RPM, TPM, RPD). RPD resets at midnight Pacific Time. RPM/TPM use a rolling window. Hard-coding quota numbers is unreliable; AI Studio shows live limits. Trying 3.6 on eve…

    minsikpaul92 · GitHub · Jul 28, 2026
  3. 3

    Implementation note — the premise of this issue was wrong in one respect, and it would have taken prod down. `gemini-3.5-flash-lite` **rejects `thinking_budget=0`** with `400 INVALID_ARGUMENT`. `GeminiProvider._build_config` (`backend/services/llm.py`) sets that unconditionally, so a pure env-var switch would have failed *every* Gemini call rather than degrading quality. Verified against the live API on 2026-08-02: | config | result on `gemini-3.5-flash-lite` | |---|---| | `thinkingBudget: 0` |…

    jrhahn · GitHub · Aug 2, 2026
  4. 4

    ## Question Which models are candidates for the Assistant's extraction call, and what does one Turn cost on each? Compare Claude Haiku 4.5 (and any cheaper current Claude tier) called directly against two or three cheap models reachable via OpenRouter. For each: price per million input and output tokens, cost of a typical Turn (about 1.5K input tokens of system prompt plus conversation, about 300 output tokens of structured JSON), structured-output or JSON-schema support, typical latency, and r…

    jdlam · GitHub · Sep 3, 2026
  5. 5

    ## Objective Find and fix why Avenger (`.github/workflows/avenger.md`) declares `model: claude-haiku-4.5` but audited runs execute on `claude-opus-4-8`. ## Issue context Issue #50312 found the largest cost driver on the Avenger workflow is a model mismatch: frontmatter/intended budget points to Haiku, but `agent_usage.json` shows Opus 4.8 in all 3 sampled runs (5x more expensive per token). This suggests a compiler/runtime bug in model propagation, alias resolution, or engine argument generatio…

    github-actions[bot] · GitHub · Aug 4, 2026
  6. 6

    ## The incident that surfaced this While pinning a cheap model for WhatsApp channel turns (chat-surface cost reduction, 2026-08-06), we set `ExecutionConfig.model_name: "composer-2.5"` and verified the pin held — but production billing records show every call resolved to **`composer-2.5-fast`**, which carries ~6x the base input/output rates ($3.00/$15.00 vs $0.50/$2.50 per M tokens). The same happened to the other calibration arm: `claude-haiku-4-5` billed as `claude-4.5-haiku-thinking`. Root c…

    whysosuresh · GitHub · Aug 6, 2026
  7. 7

    $20 cursor pro plan only, grok 4.6 is good enough and cheap

    armadillo1354 · Hacker News · Aug 14, 2026
  8. 8

    Artificial Analysis shows Grok 4.6 taking $1,068 to run their suite while Gemini 3.7 Flash takes $485. So it looks like Gemini 3.7 Flash is less than half the price in the real world. Per-token cost isn't a great metric given that some use way more tokens than others.

    mdasen · Hacker News · Aug 13, 2026
  9. 9

    ## What Forge now cost-routes **Dependabot patch/minor** PR reviews to a cheaper model (`claude-sonnet-4-5`), keeping majors and everything else on the fleet default (`claude-fable-5`). The routing decision is made **server-side** in Forge (`/api/v1/reviewer-models`, shipped in TrueTime-Health/forge#176) — but it can only route when the calling workflow hands Forge the PR's author and title. This repo's `claude-code-review.yml` doesn't yet, so its Dependabot PRs still bill the expensive reviewe…

    truetime-helper[bot] · GitHub · Aug 20, 2026

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