October 2026. The last hundred days shipped more new models than some years manage in twelve months, and the boards that rank them changed calibers mid-quarter. This is the first of a planned quarterly series on the model field: every ranking below carries its source and its snapshot date, prices are per million tokens unless noted, and anything I could not trace to a dated source stayed out. As of the October 11, 2026 data cutoff, that leaves a quarter with three Anthropic releases, a two-board disagreement at the top of text, an open-source takeover priced in dimes, and a fresh class of image, video, music, and voice models. The next snapshot is scheduled for January 15, 2027.
Two boards, two answers, one rule
Start with the mess, because everything else this quarter happens in its shadow. The Artificial Analysis Intelligence Index reached v4.3, and on the September 30 snapshot (relayed by datalearner) the #1 spot belongs to Claude Opus 5.5. A separate relay of an AA leaderboard snapshot via llm-stats, dated October 9, puts GPT-5.6 Sol at the top at 0.590 on its normalized scale. Two sources, two scales, snapshots nine days apart, two different models in first place.
The rule for this series: ranks from different calibers never merge. Each claim below cites its own board and its own date, and any article that crowns "the #1 model" without either is noise.
So who leads text? Depends which board you read and when it was snapshotted. One-line answer: as of this writing it is a two-model photo finish, Opus 5.5 on the v4.3 index (September 30), GPT-5.6 Sol on the October 9 relay, with no dated source that settles it between them.
GPT-5.6 in three lanes
The October 9 relay prices OpenAI's family in three lanes, and the spread is the story.
| Tier | Score (llm-stats relay, 2026-10-09) | Price in/out per MTok | |------|-------------------------------------|------------------------| | GPT-5.6 Sol | 0.590 | $5 / $30 | | GPT-5.6 Terra | 0.550 | $2 / $12 | | GPT-5.6 Luna | 0.510 | $0.20 / $1.20 |
Scores are normalized to that relay's own scale, so do not compare them against v4.3 index numbers anywhere else in this article. What travels between the two is the ordering inside the family: Sol first, Terra 0.040 behind, Luna 0.080 behind the flagship, at 25 times less per input token. Pay Sol prices for Luna jobs and the ledger notices before you do. A 0.080 score gap on a 0-1 board is the smallest premium I have seen justify a 25x price spread, which is a polite way of saying the cheap tier exists for a reason.
Anthropic's fifteen days
Between September 22 and October 7, Anthropic shipped three models (all dates from the Anthropic Newsroom). Claude Opus 5.5 arrived September 22 and reaches Fable 5.1's level on most work while running about 40% cheaper than Opus 5. Claude Sonnet 5.5 followed September 28 at 30% faster than Sonnet 5 and, for most work, 30% cheaper. Claude Haiku 5.5 closed the relay on October 7 as the fastest and cheapest small model in the lineup. Three tiers refreshed in fifteen days is less a launch than a handoff.
Read the vendor claims at the right altitude: each figure is a delta against Anthropic's own previous model, and the Fable 5.1 comparison for Opus 5.5 is the vendor's scope, not an independent board's. Those are still useful numbers, they just answer "better than what came before," and Anthropic is the one holding the ruler.
The open-source board, repossessed
September 22, per Economic Information Daily's report on the AA open-source board: Xiaomi's MiMo-V2.6 family took open-source #1 and domestic #1 — the flagship MiMo-V2.6-Pro scored 46 points at $0.13 per task, the first open-source entrant inside the ten-cent range, passing Kimi K3 and Qwen3.8 Max, and sits third among open models on Code Arena, level with Claude Fable 5's high-spec variant. Sina Finance's and Beijing Daily's September 22-23 syndications confirm the same figures.
The same snapshot carries the rest of the budget field: GLM-5.3-Flash from Zhipu (320B parameters, $0.15/$0.50 per million tokens), Gemini 3.7 Flash ($0.75/$3.75), Grok 4.5 from xAI ($2/$6), MiniMax M2.7 (205K context, $0.30/$1.20), and Inkling-Small from Thinking Machines Lab (276B parameters). Three of those five list input prices under a dollar. Whatever your text bill looked like in July, the floor moved.
One calibration note: 46 points on an open-source-specific board is reported here exactly as sourced and not converted onto any other board's scale. Third caliber, third date, same rule.
Images: 2.1 shipped, and the podium did not move
Nano Banana 2.1 (gemini-nano-banana-2.1) replaced the retired gemini-3.1-flash-image on October 6, per Google's changelog. The feature list is long: native 4K output, masked local repairs, subject consistency across multi-turn edits, and Chinese text rendering that improved enough for QbitAI (October 10) and Sina Tech to both flag it. Pricing moved too, roughly halved at 1K/2K and down about 25% at 4K. The model runs on Gemini 3.6 Flash with a minimal/medium/high thinking dial that defaults to medium.
Now the part the changelog will not headline. On the Arena boards it entered fourth in text-to-image, up 80 points over Nano Banana 2, fifth in multi-image editing, sixth in single-image editing. Fourth, fifth, sixth. Better model, no podium. Google keeps Nano Banana Pro for the heavy end — dense infographics, high-fidelity prototypes, complex graphic design — so the 2.1 line is a workhorse upgrade, and the boards say the workhorse still has three models ahead of it in text-to-image.
Video: a deep roster and one name that is still a rumor
Shipping now: Kling 3.0 and Kling O3 Pro, Veo 3.1 and 3.1 Fast, Runway Gen-4.5, and LTX 2 Pro. Retired this year: Sora 2 and Sora 2 Pro — the app went dark April 26 and the API followed on September 24, per OpenAI's own notices. The arrivals bin holds Seedance 2.5 with a dedicated Seedance 2.5 Video Edit variant, Wan 2.5, WAN 3.0, MiniMax H3 and H3 Max, Gemini Omni Flash and 1.1, Flux 3 Video, and HeyGen Avatar 4.
One rumor gets handled with tongs. Kling 4.0 is reportedly near launch — WaveSpeed published a launch-day testing guide on September 27 — and reported is the operative word. A launch-day test guide is not a launch. When 4.0 actually ships it earns its own article; until then it stays in this paragraph, and everything upstream of this sentence is third-party commentary about unreleased software.
Music: Suno split into three
Suno shipped the V6 family on September 9 (coverage via CSDN and others): v6 as the controllable, consistency-focused flagship, v6-wild for exploration, v6-mini free and lightweight. The edit features are where the leverage sits — natural-language local editing and single-line lyric edits turn "make the chorus less busy" into a prompt instead of a full regeneration — with multimodal input and stem separation rounding out the list, plus a hardening pass on compliance. A free lightweight tier alongside a flagship is also a quiet statement about where casual music generation now sits: one tier up from a toy.
Voice: two GAs and a pivot
September 22 took Gemini 3.8 Flash TTS and Flash-Lite TTS to GA, replacing the gemini-3.1-flash-tts-preview: studio-grade fidelity, regional dialects, long-conversation stability, voice design, consent-verified voice cloning, and a library past 150 voices. ElevenLabs answered by widening its platform toward video aggregation and shipping ElevenMusic. Voice keeps commoditizing upward — capabilities that were enterprise quotes a year ago now ship as GA line items with a consent checkbox attached.
What a quarter like this does to a platform
Since this is the paragraph where my employer enters the frame: I evaluate models for x-rush.ai, and the platform's answer to a quarter like this one is to not answer. Two names above are already wired in — Image Studio renders with Nano Banana 2.1, and Music Gen runs on Suno V6. Everything else in this article is commentary from the outside, and the product makes no promise about any particular engine. You describe the job; each task routes to a model that fits it; providers stay out of the product vocabulary and the credit price stays flat per output type whether the boards reshuffled yesterday or not. When v4.4 lands and the crown moves again, the product does not change and neither does your bill.
That design earns its keep exactly in quarters like this one. A platform that promised you "Opus" in July would owe you an apology paragraph in October.
The quarter, compressed
The quarter's real product was churn: three Anthropic refreshes in fifteen days, a two-caliber stalemate at the top of text with Opus 5.5 (v4.3, September 30) against GPT-5.6 Sol (October 9 relay), a 46-point open-source takeover by Xiaomi at ten cents a task, an image model that improved everything and topped nothing, a Suno split three ways, and TTS that finally went GA. Rankings dated above; every one of them is due to move.
Read the boards with their dates attached.
Or skip the vigil and let each task pick its own model.