Mid-range is the production spend band: expensive enough that you should measure it, cheap enough that it can carry daily work if the model fits.

Use this page as a routing guide, not a leaderboard. The right answer depends on the tool path, data policy, context size, and cost per successful task.

Current Shortlist

LanePricing anchorBest useCaveat
Claude Sonnet 5$2 / $10 current provider anchorFirst Claude cost/performance testNew tokenizer can produce about 30% more tokens for the same text
GLM-5.2$1.40 / $4.40 per 1MJuly value pick to test for long-context supported-tool codingAA and Z.AI benchmark signals still need repo-level checks
Kimi K3 API$0.30 cache-hit / $3.00 cache-miss input, $15.00 outputLatest Kimi flagship and 1M-context eval laneWeights released under the Kimi K3 License; check serving terms; CAR is not verified
Kimi K2.7 Code API$0.19 cache-hit / $0.95 cache-miss input, $4.00 outputCheaper routine Kimi coding API lane256K context; thinking is required; HighSpeed costs more
MiniMax M3 Standard$0.60 / $2.40 per 1M up to 512K inputValue coding-agent lane to testStandard vs Priority pricing and Token Plan quota are separate
GPT-6 LunaSee current guideOpenAI-native bounded work and high-volume evaluationSeparate API pricing, plan entitlements, retries, and review cost
Kimi K2.6 API$0.16 cache-hit / $0.95 cache-miss input, $4.00 outputProduct/integration-specific compatibility laneKeep only where non-thinking mode or an existing product contract still requires K2.6
Gemini Flash/Pro lanesmodel-specificHigh-context Google workflow testsVerify the exact Gemini model, tier, and data terms

How To Choose

Start With The Tool Fit

If your team lives in Claude Code or claude.ai, start with Sonnet 5 and escalate to Opus 5.5 only when Sonnet misses something important.

If your workflow supports GLM, MiniMax, Kimi, Gemini, or GPT-6 through official or BYOK routes, run the same bug fix, refactor, and review task through those lanes before paying premium-model rates. GLM-5.2 remains a value pick to test because it combines 1M context, published API pricing, independent evidence, and supported-tool subscription routing. Kimi K3 is the current Kimi flagship to evaluate; its output price still makes CAR testing mandatory. Use GLM-5.2 vs Kimi K2.6/K2.7 for the cheaper Z.AI vs Kimi decision.

Measure Cost Per Success

Token price is only part of the bill. Track:

  • prompt and output tokens,
  • retries,
  • failed patches,
  • human review time,
  • tests passed on first attempt,
  • data policy and retention fit.

Escalate Deliberately

Move to Premium Models when the task proves it needs a stronger model:

  • architecture or multi-file reasoning fails,
  • long-horizon agent work loops,
  • code review misses expensive defects,
  • safety or compliance review needs a second pass,
  • OpenAI or Claude tooling fit matters more than raw token price.

Production Routing Pattern

If the task is…Start withEscalate to
Routine coding or summarizationSonnet 5, GLM-5.2, or Kimi K2.7 CodeKimi K3 or Opus 5.5 if quality matters
Supported-tool budget codingGLM-5.2 or GPT-6 LunaGPT-6 Sol or Opus 5.5 when the value lane misses
Long-context agent workMiniMax M3 or Gemini laneOpus 5.5, then revisit Fable 5 only if access, compliance, and guardrail tests pass
OpenAI-native workGPT-6 Luna via the premium guideGPT-6 Sol when the harder tier changes the result
Guardrail-sensitive cyber/bio/chem workApproved access path firstAvoid Fable unless fallback/refusal is acceptable

Sources

GPT-6 and Opus 5.5 pointers were refreshed September 27, 2026. GLM, Kimi, MiniMax, Gemini, and other provider evidence retain their source dates; verify live pricing, availability, and open-weight status before committing large workloads.