Model Selection for Coding Tasks

This is a maintained, opinionated guide to choosing models for coding, review, and agentic work; prices and benchmarks below are as of 2026-10-04.

Model selection for coding is an economics decision before it is a capability decision: run the workhorse tier for the loop, buy the frontier by the minute, and give the reading to the cheap models. The part most engineers get backwards: for anyone paying per token, the harness you run moves your bill about as much as the model you pick, and sometimes more.

Task class decides the tier

#

Three task classes cover almost all coding work, and each maps to a price tier, not to a leaderboard rank.

• Completion and small edits: the model writes a few dozen lines from tight context; latency and price matter more than peak reasoning, so the cheap tier wins (gpt-6-luna at half its predecessor’s rate, Gemini 3.1 Flash-Lite, Claude Haiku 4.5, GLM-5.3-Flash).
• The agentic loop: explore, edit, test, repeat for minutes or hours; this is where most tokens die, and the mid tier (Claude Sonnet 5.5, gpt-6-sol, Gemini 3.1 Pro Preview) resolves most of it, while the Kimi, GLM, and DeepSeek challengers below price the same class far lower.
• Hard planning and review judgment: architecture choices, gnarly debugging, deciding what to let through; this is the only class where frontier spend reliably pays, and it is a small fraction of your turns.

Escalate by task class inside a session instead of picking one model for everything; Amp’s low/medium/high/ultra modes and Codex’s Sol/Terra/Luna defaults are this idea shipped as product (see the Amp and Codex notes).

The lineup as of 2026-10-04

#

The workhorse tier has converged to roughly $2 in and $10-12 out per million tokens at every major provider, which means switching costs are now measured in harness integration, not price. GPT-6 Sol joined the club on 2026-09-22 at exactly $2/$10, the first OpenAI model to price at the converged rate, while gpt-5.6-sol stays at $4/$20 behind its promotion. On 2026-09-29 OpenAI refreshed the tier again with gpt-6.1-sol at the same $2/$10 and the cache hit halved to $0.10. Claude added Opus 5.5 the same day at $4/$20 with a 5 percent cache-hit rate, the first Claude tier priced between Sonnet and Opus, and Claude added Sonnet 5.5 on 2026-09-28 at the unchanged $2/$10 workhorse rate with the standard one-tenth cache hit. gpt-5.6-terra and gpt-5.6-luna keep flipping on and off OpenAI’s pricing page, off on 2026-09-27, back on 2026-09-29, off again earlier on 2026-10-04, and back on at launch prices by my second look the same day, with OpenRouter serving both at those rates throughout, OpenAI’s Daybreak line still carries gpt-5.6-cyber at $12.50/$75, and Gemini’s 3.8 Flash Cyber sells at $1.50/$7.50, the standard rate the regular Flash family only moves to on 2027-01-01.

A ? marks a property the verified pricing pages do not state. The lineup keeps only models released within the last year, the oldest entry is Claude Haiku 4.5 on 2025-10-15, 11 days inside the window, and release dates and context windows follow the models.dev list reference, which also supplies the newest Gemini Flash siblings the table groups on shared intro pricing, and deepseek-flash’s 2026-09-10 release date now that models.dev has caught up.

Four details the table hides:

• Batch mode is 50% off at OpenAI, Anthropic, Google, and Qwen3.8-Max, which makes overnight review sweeps half price by default.
• Cached reads run about one tenth of input price at OpenAI, Google, and Kimi K3, one tenth at Anthropic except Opus 5.5 at one twentieth and the Fable 5.1 and Mythos 5.1 pair at one fortieth, closer to one fifth at GLM and on kimi-k2.7-code, Qwen discounts hits without publishing a rate, and DeepSeek bills hits at about one fiftieth on Flash and one thirtieth on V4 Pro, so loop economics depend on both the cache-hit rate and the vendor’s cache discount.
• Claude models from 4.6 onward include the full 1M-token context at standard pricing, the cheapest flat-1M input prices belong to the challengers (deepseek-flash at $0.30 peak, GLM-5.3 at $1.40, Qwen3.8-Max at $2, Kimi K3 at $3), and prices double past the long-context thresholds at OpenAI and on Gemini 3.1 Pro.
• The Claude 4.7+ tokenizer produces about 30% more tokens for the same text, so cross-vendor price comparisons understate Anthropic’s effective cost by roughly that margin.

The challengers reset the price floor

#

Kimi, GLM, DeepSeek, and Qwen price the agentic loop at half the converged workhorse rate or less, which makes the $2/$10-12 “standard” a choice rather than a fact. GLM-5.3, 5.2, and 5.1 sit at $1.40 in and $4.40 out per million tokens, with GLM-5 at $1/$3.20, per Z.ai’s pricing page as of 2026-09-27. Kimi’s coding-specialized kimi-k2.7-code lists at $0.95/$4.00 with a 256K context and multimodal input, and Moonshot’s own pricing and its official OpenRouter endpoint agree on that price as of 2026-09-25. Third-party OpenRouter hosts serve the same model id from $0.67/$3.35, but that floor is hosts undercutting each other weekly, not a Moonshot list price. Its HighSpeed variant lists at $1.90/$8.00 for about 180-260 tokens per second of output. Kimi K3 takes the other flank: $3/$15 list, which Moonshot’s page and its official endpoint both state, always reasoning with a configurable effort, and a flat-price 1M context that only Claude otherwise offers, with the same host floor running from $0.67/$13.00 for token-payers who accept cheap-host variance. Z.ai even keeps GLM-4.7-Flash free, which makes it the zero-dollar candidate for inline completion and routing experiments. GLM-5.3-Flash fills the gap between free and $1.40: its 50 percent launch promo expired at 24:00 on 2026-09-09 Singapore time exactly as scheduled, and Z.ai’s page billed the $0.15/$0.50 list price with cached reads at $0.03 when I re-checked on 2026-09-29. On 2026-09-18 Z.ai wedged a new rung between Flash and GLM-5.3: GLM-5.3-FlashX, a high-speed Flash serving option for coding and agent workflows at $0.37/$1.25 with cached reads at $0.075. DeepSeek V4 undercuts them all: v4-pro lists at $1.32 in and $3.96 out per million at peak, with every hour outside 01:00-04:00 and 06:00-10:00 UTC on weekdays billed at half, so off-peak runs pay $0.66/$1.98. The V4 Pro retirement announced on 2026-09-10 is off: DeepSeek’s pricing page said on 2026-09-12 that V4 Pro service continues past 2026-09-14 with billing unchanged, so this row is permanent again. On 2026-09-10 DeepSeek replaced v4-flash outright: deepseek-flash (V4.1-Flash) lists at $0.30/$1.20 peak and $0.15/$0.60 off-peak, adds native vision input, and both models keep a flat 1M context with 384K max output and thinking mode on by default. DeepSeek’s cache hits bill at $0.044 (pro) and $0.006 peak or $0.003 off-peak (flash) per million, the Flash hit about one fiftieth of a miss, the deepest cache discount of any vendor in this guide. It also exposes an Anthropic-format endpoint, so harnesses that speak the Claude API can point at DeepSeek directly. Qwen3.8-Max rounds out the field at $2 in and $6 out per million tokens with flat 1M pricing, at or below the converged workhorse rate at flagship quality, and Qwen3.8-Flash sits in the cheap tier at $0.15/$0.47. Alibaba’s Model Studio is not a native OpenCode provider the way Moonshot, Z.AI, and DeepSeek are, but it speaks the OpenAI-compatible format, so Qwen rides OpenRouter or any compatible endpoint. Harness fit is solved: OpenCode lists Moonshot AI, Z.AI, and DeepSeek as native providers, so the lean BYOK harness plus a challenger model is one /connect away. My disagreeable claim: for a token-payer, the rational default loop in September 2026 is deepseek-flash, kimi-k2.7-code, or GLM-5.3 through OpenCode, and the big-three workhorses are what you escalate to, not what you default to. Whether the challengers hold quality on your codebase is exactly what SWE-bench’s bash-only view and a two-week cost-per-merged-PR measurement are for; the economics alone no longer justify defaulting to the big three. The era of Gemini Flash’s $0.75 intro rate leading a price category is over: gpt-6-luna bills $0.10/$0.50, GLM-5.3-Flash matches its output rate at $0.15/$0.50, and deepseek-flash beats both on input at $0.30 peak, $0.15 off-peak, with no expiry announced.

Benchmarks decay faster than prices

#

Benchmark points do not track price, and the last independent leaderboard that published cost per run makes the point brutally. The aider polyglot leaderboard froze in 2025 along with the tool’s maintenance (see the aider note), but its dated runs still teach the right lesson. In October 2025, DeepSeek-V3.2-Exp (Reasoner) scored 74.2% for $1.30 of API spend across the 225-exercise suite. In June 2025, o3-pro (high) scored 84.9% for $146.32: ten points more for 113 times the money.

For current rankings, SWE-bench Verified is the reference, and its default view now runs every model in the same minimal mini-SWE-agent bash environment precisely so harnesses stop inflating scores. The frontier of evaluation moved on again: CodeClash (November 2025) scores goal-oriented development and ProgramBench (May 2026) scores building artifacts from scratch. My rule: use benchmarks for relative history and cost-per-point for decisions, and never quote a vendor’s self-reported number, including the SWE-Rebench claim on Junie’s own site (see the Junie note).

The harness multiplies the price of everything

#

The measured spread between harnesses is larger than the frontier premium at OpenAI and most of it at Anthropic. The July 2026 proxy study recorded in the Claude Code note measured about 33k input tokens sent before the user’s prompt on a minimal task, mid-session cache re-writes up to 54x, and a 4.2x multiplier on a two-subagent fan-out. The same study, recorded in the OpenCode note, measured about 7k baseline tokens for OpenCode with byte-stable cache prefixes, making it about 3.7x cheaper on a matched pass/fail benchmark. Compare: Fable 5.1 costs 5x Sonnet 5 per token, and gpt-5.6-cyber costs more than 6x gpt-6-sol. So choosing OpenCode over Claude Code moves a token-payer’s bill more than choosing Sonnet 5 over Opus 5.5, and roughly matches skipping Fable 5.1 for Sonnet 5. Models are substitutable across providers; a harness with a heavy baseline taxes every model you point through it, forever.

Subscription or BYOK: pick by where your spend lands

#

Subscriptions win below a spend threshold and handcuff you above it, so the unit that decides is dollars of monthly token spend, not features. The subscription side: Claude Code is included from Claude Pro ($20 monthly) through Max 20x ($200/month) per the Claude Code note, and Codex is included from the ChatGPT Free tier upward per the Codex note, though usage meters in shared five-hour windows with wide published ranges (10-100 local Sol messages on Plus). Amp’s Megawatt at $20/month makes a linked ChatGPT subscription’s usage effectively unlimited at low and medium modes per the Amp note, the cheapest delegation entry if you already pay for ChatGPT. The BYOK side: OpenCode, Junie, Crush, and aider bill at provider rates with zero markup (Junie’s cloud credits, 10 per 30 days on AI Pro, make BYOK effectively mandatory for real work per the Junie note). The crossover math: $100 of Sonnet 5 API spend is 50M raw input tokens at the table’s $2/M rate, and closer to 20M total tokens once you blend in output-priced generation at agentic ratios, which one heavy week can eat; if you reliably run past that, subscriptions cap your downside until their limits bite. One fence to remember: Anthropic’s March 2026 legal requests removed Claude subscription login from OpenCode (see the OpenCode note), so subscription value does not transfer to third-party harnesses; subscribe to the vendor whose harness you will actually run.

A decision guide

#

Everything above collapses into one sentence per situation, and every price claim in it is dated in the table above.

• Daily loop default: Claude Sonnet 5.5, gpt-6-sol, or Gemini 3.1 Pro Preview; pick by harness fit, they price the same, and Kimi K3 at $3/$15 official (the OpenRouter host floor is roughly a fifth of official input, with cheap-host caveats) is the token-payer’s challenger option in that class.
• High-volume loops where cost leads: deepseek-flash at $0.30 peak and $0.15 off-peak with no expiry announced, gpt-6-luna at $0.10/$0.50, GLM-5.3-Flash at $0.15/$0.50, or Gemini 3.8/3.7/3.6 Flash at the shared $0.75/$3.75 intro rate (calendar the 2026-12-31 end).
• Price-floor loops, BYOK: deepseek-flash, kimi-k2.7-code, or GLM-5.3 through OpenCode, all native providers; nothing credible stays cheaper.
• Whole-repository reads: DeepSeek from $0.30 in, a 1M-context Claude or Qwen3.8-Max at $2, or Kimi K3 at $3 flat (host floor $0.67); all four keep 1M pricing flat.
• Frontier minutes only: gpt-6-astra, gpt-5.6-sol, Claude Opus 5.5, or Claude Opus 5 for planning and stuck debugging; Fable 5.1 when nothing else resolves.
• Inline completion and edits: gpt-6-luna, Gemini 3.1 Flash-Lite, Claude Haiku 4.5, or GLM-5.3-Flash.
• Review passes and doc reading: cheap tier in batch mode, or a 1M-context Claude for whole-repo reads at standard price.
• Offline or air-gapped: local models through aider, Crush, or Junie BYOK.
• Already paying ChatGPT: run Codex, and consider Amp for delegated work.
• Already paying Claude: run Claude Code; do not expect that subscription to drive any other harness.

What changes fast and how to re-verify

#

Everything volatile in this guide is dated, and four clocks are running. GLM-5.3-Flash’s half-price launch promo expired at 24:00 on 2026-09-09 Singapore time as scheduled, and Z.ai’s page billed the $0.15/$0.50 list price when I re-checked on 2026-09-29. DeepSeek cancelled the planned 2026-09-14 routing of v4-pro requests to V4.1-Flash and V4 Pro keeps billing at $1.32/$3.96 peak, Sol’s promotional pricing runs at least through 2026-11-21, the Gemini Flash intro rate ends 2026-12-31, and benchmark relevance decays on roughly a quarterly cycle (CodeClash, then ProgramBench within six months). The fourth clock is impersonal: OpenRouter’s third-party host floors on the Kimi models moved again this week (K3’s floor is now $0.67/$13.00 on Relace, with InferenceNet at $0.72/$13.00 and the aggregator listing tracking $0.72/$13.00), so treat those floors as spot prices and anchor purchasing decisions to the list prices. Sonnet 5 shows the other direction: a scheduled September 2026 increase to $3/$15 was cancelled weeks before taking effect, so scheduled changes are announcements, not facts. On each refresh I re-fetch the seven provider pricing pages (OpenAI, Anthropic, Google, Moonshot, Z.ai, DeepSeek, Alibaba) plus OpenRouter’s models and endpoints APIs and the models.dev list, the SWE-bench leaderboards, and one current harness-overhead measurement, and I update the table above and the as-of date together. A fact that cannot survive that re-fetch gets deleted rather than hedged.

What to Do Next

#

• Set your harness default to the workhorse tier this week, and escalate to a frontier model only for planning steps and stuck loops.
• Measure cost per merged PR on your own repository for two weeks; that number outranks every benchmark cited here.
• If you pay per token, prefer a harness with byte-stable cache prefixes, because the 54x cache re-write failure mode costs more than any model choice.
• Route review sweeps through batch mode; half price at all four providers that offer it is the cheapest quality improvement available.
• The 2026-09-10 GLM-5.3-Flash check resolved as predicted: the promo expired into the $0.15/$0.50 list price; calendar the 2026-11-21 and 2026-12-31 promo deadlines to re-run your economics then.
• If you point a loop at DeepSeek, schedule the long runs off-peak; half price by the clock is the deepest vendor discount in this guide and it stacks with caching.

Changes

#

• 2026-08-24 - Created as the queue essay on task-class tiers; added the Kimi, GLM, and DeepSeek V4 lineups the same day on owner request.
• 2026-08-26 - Clarified the $100 crossover math so it stopped reading against its own table and added DeepSeek to the challengers mention.
• 2026-09-04 - Folded Claude Fable 5.1 into the lineup table and updated the cache bullet for its reduced cache-hit rate.
• 2026-09-08 - Transposed the lineup table on owner request so models read as columns against attribute rows, values unchanged.
• 2026-09-09 - Recorded the kimi-k2.7-code price drop to $0.71/$3.50 on OpenRouter and replaced the HighSpeed 2x claim with dollar figures.
• 2026-09-10 - Added DeepSeek’s deepseek-flash pricing, retired V4-Flash, and added the 2026-09-14 v4-pro routing clock.
• 2026-09-12 - Recorded kimi-k3’s OpenRouter drop to $2.65/$13.28 and removed the cancelled Flash-routing clock.
• 2026-09-16 - Recorded OpenRouter’s kimi-k2.7-code cut to $0.71/$3.21 with cache hits at $0.18 while Moonshot’s own page still lists $0.95/$4.00.
• 2026-09-18 - Recorded OpenRouter’s kimi-k3 cut to $2.10/$10.95 with cache hits at $0.23 while Moonshot’s own page still lists $3/$15.
• 2026-09-20 - Recorded OpenRouter’s kimi-k3 cut to $1.70/$8.50 with cache hits at $0.17 while Moonshot’s own page still lists $3/$15, and added Z.ai’s GLM-5.3-FlashX, released 2026-09-18 at $0.37/$1.25, to the lineup table and the challengers section.
• 2026-09-24 - Replaced the verification preamble with a single as-of clause in the intro on owner request.
• 2026-09-25 - Added GPT-6 Sol ($2/$10) and GPT-6 Luna ($0.10/$0.50), released 2026-09-22 per models.dev and launched to a 1,762-point thread, to the lineup table; the workhorse paragraph, cheap-tier bullets, and the price-category sentence updated.
• 2026-09-25 - Re-grounded the Kimi rows on Moonshot’s official prices after OpenRouter’s endpoints API showed the previously recorded cuts were third-party host floors, not vendor list prices: kimi-k2.7-code back to $0.95/$4.00 (cache $0.19) and kimi-k3 to $3/$15 (cache $0.30), with the host floors named in prose and the re-verify list now including the endpoints API.
• 2026-09-25 - Corrected the GLM-5.3-FlashX cached-read cell from $0.075 to the $0.090 Z.AI serves, and fixed the context-behavior row where the gpt-6-astra cell (doubles past threshold) and the Haiku 4.5 cell (200K) had been transposed.
• 2026-09-26 - Linked the new Model Access Feature Matrix, the layer of gateways, coding plans, and flat subscriptions that packages the per-token prices this guide tracks.
• 2026-09-27 - Refreshed every provider price: added Claude Opus 5.5 (2026-09-22, $4/$20, 5% cache) as a column, recorded gpt-5.6-terra and gpt-5.6-luna leaving OpenAI’s pricing page plus the new gpt-5.6-cyber and Gemini 3.8 Flash Cyber, corrected GLM-5.3-FlashX’s cached read back to $0.075, moved the Kimi K3 host floor to $0.88/$4.90, fixed the context-behavior row (Haiku 4.5 to 200K, Sonnet 5 and the Gemini Flash group to 1M flat), and corrected the re-verify list to seven provider pages including Alibaba.
• 2026-09-29 - Added Claude Sonnet 5.5 (2026-09-28, $2/$10, standard one-tenth cache) as a column, corrected the gpt-5.6-terra and gpt-5.6-luna cells after both returned to OpenAI’s pricing page at their launch prices, and moved the Kimi K3 OpenRouter host floor to $0.40/$10.00 (Relace).
• 2026-10-02 - Added gpt-6.1-sol (released 2026-09-29 per models.dev, $2/$10 with the cache hit halved to $0.10) as a column, re-verified all seven provider pricing pages with zero other vendor moves, and moved the OpenRouter host floors to their current spot readings (kimi-k2.7-code $0.67/$3.35, K3 $0.38/$10.00 on Relace).
• 2026-10-02 - Moved the Kimi K3 OpenRouter host floor from $0.38/$10.00 (Relace) to $1.44/$14.00 (Wafer) and the aggregator listing to $2.70/$13.50 within the same day; Moonshot’s official $3/$15 and every other provider anchor re-verified unchanged.
• 2026-10-03 - Moved the Kimi K3 OpenRouter host floor again, from $1.44/$14.00 (Wafer) to $0.98/$13.00 (Relace) with InferenceNet at $0.99/$13.00; Moonshot’s official $3/$15, the aggregator listing ($2.70/$13.50), the k2.7-code floor ($0.67/$3.35), and all seven provider anchors re-verified unchanged.
• 2026-10-04 - Recorded gpt-5.6-terra and gpt-5.6-luna dropping off OpenAI’s pricing page a second time (OpenRouter still serves both at launch rates) and re-anchored the as-of date to today.
• 2026-10-04 - Moved the Kimi K3 OpenRouter host floor again, from $0.98/$13.00 (Relace) to $0.40/$13.00 (Relace) with InferenceNet at $0.72/$13.00 and the aggregator listing at $0.50/$13.00; Moonshot’s official $3/$15 (via its pricing page and endpoint), the k2.7-code floor ($0.67/$3.35), and every other provider anchor re-verified unchanged against all seven pricing pages, OpenRouter, and models.dev.

See also

#

• Model Provider Feature Matrix - this guide’s prices regrouped into a provider-by-provider bundle comparison
• Model Access Feature Matrix - the gateways, coding plans, and flat subscriptions you can buy these models through
• Agentic Coding Tools Landscape - the four-layer map this guide’s harness claims sit inside
• Claude Code - the measured token overhead that anchors the harness-multiplier argument
• Codex - the Sol/Terra/Luna tiering as shipped defaults
• OpenCode - the lean BYOK counterexample and the subscription-transfer fence
• aider - the frozen leaderboard whose cost data this guide leans on

References

#

• https://platform.openai.com/docs/pricing - GPT-5.6 and GPT-6 family token prices including the 2026-09-29 gpt-6.1-sol ($2/$10, $0.10 cache read), batch discount, promo end date, long-context doubling; gpt-5.6-terra and gpt-5.6-luna back on the page at launch prices as of this fetch, after dropping off earlier the same day (fetched 2026-10-04)
• https://platform.claude.com/docs/en/about-claude/pricing - Claude 5 lineup including the 2026-09-28 Sonnet 5.5 ($2/$10), cache multipliers, 1M-context policy, tokenizer note, Sonnet 5 price decision (fetched 2026-10-04; the docs.claude.com URL redirects here)
• https://cloud.google.com/vertex-ai/generative-ai/pricing - Gemini 3 family pricing, intro windows, long-context and batch rates (fetched 2026-10-04)
• https://www.swebench.com/verified.html - the Verified benchmark and the bash-only mini-SWE-agent comparison setup (fetched 2026-10-04)
• https://www.swebench.com/ - CodeClash and ProgramBench launch dates, mini-SWE-agent 65% result (fetched 2026-10-04)
• https://aider.chat/docs/leaderboards/ - the polyglot leaderboard with per-run costs and dates used for the cost-per-point argument (fetched 2026-10-04, still frozen at its 2025 state)
• https://platform.kimi.ai/docs/pricing/chat-k3 - Kimi K3 prices ($3/$15), cache-hit rate, flat 1M context, always-on reasoning with configurable effort (fetched 2026-10-04; the page renders its tables through a JS shell, so I read the same URL’s markdown render plus Moonshot’s official OpenRouter endpoint)
• https://platform.kimi.ai/docs/pricing/chat-k27-code - kimi-k2.7-code’s official prices ($0.95/$4.00, matching Moonshot’s own OpenRouter endpoint re-checked 2026-10-04), 256K context, HighSpeed variant prices and speeds (fetched 2026-10-04, same JS-shell note)
• https://docs.z.ai/guides/overview/pricing - GLM-5.x family prices, cached input rates, free Flash tiers, the GLM-5.3-Flash list price billed after the 2026-09-09 promo, and GLM-5.3-FlashX’s 2026-09-18 price with its cache read (fetched 2026-10-04; all cells unchanged)
• https://opencode.ai/docs/providers/ - Moonshot AI, Z.AI, and DeepSeek as native OpenCode providers
• https://api-docs.deepseek.com/quick_start/pricing/ - DeepSeek lineup including V4.1-Flash and the notice that V4 Pro service continues past 2026-09-14 with billing unchanged, peak and off-peak rates, 1M context and 384K output, cache-hit prices, peak-hour definition (fetched 2026-10-04, all cells unchanged; the trailing slash matters: without it the URL serves the first-API-call page)
• https://api-docs.deepseek.com/news/news260910 - the DeepSeek-V4.1-Flash release announcement of 2026-09-10: model id, V4-Flash retirement, the original 2026-09-14 V4 Pro routing plan since superseded by the pricing page’s continuation notice, and open weights
• https://help.aliyun.com/zh/model-studio/billing-for-model-studio - Qwen3.8-Max and Qwen3.8-Flash official per-token prices, batch half price on Max, context-cache discount, the 0-1M token tier, and the new-account free quota (fetched 2026-10-04, price structure unchanged)
• https://openrouter.ai/api/v1/models - aggregator listings for qwen3.8-max-0902 ($2/$6), qwen3.8-flash ($0.15/$0.47), the GPT-6 pair, and gpt-6.1-sol ($2/$10), re-checked 2026-10-04; kimi-k2.7-code’s listing there is the third-party host floor ($0.67/$3.35), not Moonshot’s list price, and kimi-k3’s listing ($0.72/$13.00) now tracks the host floor
• https://openrouter.ai/api/v1/models/moonshotai/kimi-k3/endpoints - per-provider endpoint prices separating Moonshot’s official $3/$15 (cache $0.30) from the host floor at $0.67/$13.00 (Relace), re-checked 2026-10-04
• https://models.dev - the community model list used as the lineup reference: model ids, release dates, and context windows for every model in the table, including the GPT-6 pair’s and Opus 5.5’s 2026-09-22 release dates, Sonnet 5.5’s 2026-09-28, and gpt-6.1-sol’s 2026-09-29 (fetched 2026-10-04)