Every price that fell this week fell for a segment; key your router on prompt length and call type, not list price

Registry movement

The decision-model category now has four confirmed hosted vendors (TypeSafe's Jev, OpenAI gpt-6-luna via the Decisions API, Microsoft-Decision-1, Cloudflare Clef and Clef-flash) plus one unverified (Perplexity's hosted decider, listed in a third-party ladder but not checked against Perplexity's pricing page), and at least two downloadable implementations (pplx-decider-v1.1-27b, the W40 llama.cpp task heads; Liquid AI's d1 pending verification); Microsoft's benchmark appendix (per its post, outside the graded set) names five more decision models not reviewed here (H2O-Lightning-4B, deck-31B, Strands-Decider 2B, Quyet-1.0-Large, Rune 26B-A4B). The independent-vendor condition of the 3-of-6 rule is met by any count. The remaining question is whether a typed-output classifier head on a shared base is an architecture or a fine-tuning recipe; Microsoft's Qwen3.5-9B disclosure is the first vendor statement on that question and points to recipe. The test will be applied in full when a vendor discloses a non-fine-tune build or a third decision model ships with a technical report.

Added

claude-haiku-5-5, mistral-large-4, embeddinggemma-2

Updated

claude-sonnet-5-5

Anthropic · 2026-10-07 · Frontier · Reasoning

Claude Haiku 5.5

Anthropic's first sub-dollar 5.5 model, with a two-tier rate card and a tokenizer that counts more

The question for a platform team is where its prompts fall relative to 100K tokens, because the price steps 5x there and the new tokenizer brings that step 20-23% sooner on the same text (around 77K tokens as the old tokenizer counted them). The rate card is in the architecture watch; Anthropic's blended estimate across its own customers' mix is about 75% cheaper (vendor figure), which is the number to test against your own token mix. Capability is the reason to care: Terminal-Bench 4.0 goes from 0.0% on Haiku 4.5 to 32.8% on Artificial Analysis's independent run at max effort (39.2% on Anthropic's harness), and the independent index at max effort is 43.4, above GPT-6 Luna (38). Safeguards are not uniform across the 5.5 line: Anthropic's launch page says Haiku 5.5's cyber safeguards are more restrictive than Haiku 4.5's but somewhat less restrictive than its recent models', while its biology safeguards match Sonnet 5, Sonnet 5.5 and Opus 5. The model drops temperature, top_p, top_k, budget_tokens and prefill, has 1M context with 128K output, and carries a retirement floor of October 7, 2027. It is also the obvious subagent and reviewer model for agent harnesses, which is the Agent Techniques angle.

Model registry ID

claude-haiku-5-5

Sources Anthropic

Mistral AI · 2026-10-06 · Frontier · Moe

Mistral Large 4

A trillion-parameter European MoE, callable today, downloadable at the end of the month under a licence not yet named

Buyers evaluating a sovereign-compute option should treat this as a preview with two moving parts: a price that doubles when the launch discount ends ($0.68 / $2.09 to $1.36 / $4.18; Mistral has not published the end date) and a licence that has not been named, with on-prem access today under a bespoke self-deployment agreement. The model is 1.05T total / 52B active with a 1.6B vision encoder, 1M context per Mistral's docs (Artificial Analysis measures the preview at 524K) and 160+ languages, trained from scratch on 3,800 Grace Blackwell GPUs in Mistral's own European data centres, one of two frontier-scale training-footprint disclosures this week (Reflection's is the other). Independent scores set the level: 38 on the Artificial Analysis index and 22.73% on Vals AI's Terminal-Bench 4.0 against Mistral's own 28.3%. Artificial Analysis and Willison's 'Mistral is back, about six months behind' frame is theirs; the addition here is that at $1.13 per index task (Artificial Analysis, list price) it is dearer than the open Chinese models it trails, so the case is residency and provenance, not price or capability.

Model registry ID

mistral-large-4

Sources Mistral AI

Microsoft · 2026-10-09 · Specialist

Microsoft-Decision-1

A post-trained Qwen3.5-9B decision model on Foundry at $0.042 per million input tokens, matching TypeSafe's Jev and 58% below OpenAI's beta

This is the datapoint that confirms OpenAI's Decisions API as the most expensive offer in its category three days after launch (eesel AI and CTOL Digital published the comparison on October 9), and it is why OpenAI is under pressure to cut at GA rather than hold price (the AI Stack Weekly puts a cut to $0.05 or below at 41%). Microsoft's post, outside the graded set, discloses more than most entrants: the base is a post-trained Qwen3.5-9B, the comparison covers 36 benchmarks and about 150,000 questions against six other decision models, and the vendor-measured P50 latency is about 35x faster than GPT-6 Sol; no independent run exists yet, so the model is deferred on the tree with the category. For architects the relevance is in the approve, flag and route call: with Jev and Microsoft at $0.042, Cloudflare reported at $0.038, Perplexity's open 27B at zero marginal licence cost and OpenAI at $0.10, the price of a verdict sits at or below four cents regardless of which cloud the agent runs in. For the weekly's small-reviewer prediction (March 31, 2027), a named sub-10B decision model at a major agent-platform vendor is a candidate that is not yet documented in Copilot's approval path.

Sources Microsoft Foundry (Command Line blog)

Reflection AI · 2026-10-05 · Open Frontier · Moe

Reflection Beam

A 501B / 23B-active coding MoE pretrained in under four weeks, announced as open-weight with nothing yet to download

Do not plan a deployment on this until the repository exists. The architecture and training disclosures are the substantive part: 23.8T tokens in under four weeks on 6,144 GB300 NVL72 GPUs, then 100M+ RL rollouts on 10.5K GB300s with about 1.3B sandboxes, which is a data point on how fast a frontier-scale pretrain now runs on a Blackwell Ultra fleet. Every score is vendor-reported, and on both agentic-coding rows its own table shares with the Chinese open models (DeepSWE v1.1, tabulated below, and Terminal Bench v2.1) Beam trails all three, which the-decoder's 'most capable open-weight model built outside China' frame does not reconcile. Access is a waitlisted API with no price; weights, model card and technical report are promised under Apache 2.0 after red-teaming. Deferred on the tree.

Sources Reflection AI

Google DeepMind · 2026-10-06 · Edge · Multimodal

EmbeddingGemma 2

A 740M Apache 2.0 embedder that puts text, code, image, video and audio in one 768-dimensional space and runs in 567 MB on a phone

Teams running retrieval on-device or at the edge get a single embedding space across four modalities plus code at a size that fits a handset: 270M text backbone plus optional 170M vision and 300M audio encoders, 8,192-token context, Matryoshka truncation to 512, 256 or 128 dimensions, about 191 MB active RAM text-only and 567 MB fully multimodal on a Pixel 11 Pro. Google reports MTEB (Code) at 78.68 against 68.76 for the first EmbeddingGemma, a 14% gain, and 61.36 on multilingual MTEB; both are vendor-run, and the community's in-browser WebGPU demos within a day are evidence of loadability rather than quality. EmbeddingGemma 1 had more than 20M downloads, which is the distribution argument. Placed on encoder_only, where the tree already carries text-embedding-3.

Model registry ID

embeddinggemma-2

Sources Google (The Keyword)

Perplexity · 2026-10-07 · Specialist · Dense

pplx-decider-v1.1-27b

The open decision model gets a provenance manifest and a 61.56 Decision Index, up from 56.4, and stays deferred as a category question

Last week's provenance problem (shards byte-identical to a community checkpoint) is answered: v1.1 ships with training code, a data manifest and a release manifest with file checksums under Apache 2.0, an independent header audit confirms the 26.09B-parameter layout, and OpenRouter listed it the same day. Perplexity reports 61.56 on its own Decision Index against 56.4 for v1 and a figure more than 3.5 points above TypeSafe's Jev; the index is Perplexity's, so treat it as a vendor benchmark. For buyers it is the only four-cent-class decision model whose weights, data lineage and licence can all be inspected, which matters more in an approval path than in a chat window. Deferred on the tree with the rest of the category.

Sources Hugging Face

Kandinsky Lab · 2026-10-06 · Specialist · Multimodal

Kandinsky 6.0 Video

A 3B Lite and 29B Pro video-plus-audio family released under MIT with eight Diffusers checkpoints

Outside the tree's scope but inside the open-weights pattern of the week: Kandinsky Lab shipped code, weights and Diffusers integration for a 5-second video generator with synchronized 44 kHz audio and lip-sync on the same days two Western labs announced language models that cannot be fetched. Eight checkpoints are on Hugging Face, including 10-step and 2-step distilled variants, and the README publishes per-GPU generation times (Pro at 1080p: 1,247 s on an RTX 4090, 402 s on an H100). For media teams the 3B Lite is the one that runs on a consumer card; no independent quality evaluation exists yet.

Sources GitHub (kandinskylab/kandinsky-6)

Price discrimination by prompt length, speed and call type

This is the fact home for the week's rate cards. Five prices moved and none of them moved uniformly across a model's demand curve: Haiku 5.5 is a tenth of Haiku 4.5 for short prompts and half of it past 100K, where the newer tokenizer (which counts the same text as roughly 1.25-1.3x the tokens, per Anthropic's migration notes and Willison's measurement) arrives sooner; Luna steps up at 272K (2x input, 1.5x output); Sol Ultrafast sells the identical model's output at six times list for a vendor-claimed 8x speed; Mistral's price is a launch discount that reverts on a date Mistral has not published; and the Decisions API meters input only, which is attractive for long-document verdicts and irrelevant for short ones. The circulating 'price war reaches the bottom tier' frame is Dbggr's (October 10) and Latent Space's (October 8); the reading here is segmentation. The operational consequence is that a router keyed on list price will overpay on two of these five (Haiku above 100K, and Sol Ultrafast on batch work where nobody is waiting) and underprovision on a third (the decision tier, where the input-only meter makes long-document verdicts cheaper than list suggests). Key it on three fields: prompt-length band, latency class (Ultrafast only where a human is waiting), and call type (typed verdicts to the decision tier). On the prompt-length band, the operator's control is the compaction window: Zvi Mowshowitz's advice to set autocompact at 100K is the origin of 'compact before the step', and on the new tokenizer the step falls 20-23% sooner, at about 77K of old-tokenizer text; one practitioner report (wmedia.es, October 9) finds a 100K window thrashes in Claude Code and 130K is the working setting, so measure before pinning. Measure your own token mix under the new tokenizer before trusting any vendor blended-saving figure; Anthropic's own blended number for Haiku is 75% (vendor-reported), not 90%.

Examples

Claude Haiku 5.5 ($0.10 / $0.50 to 100K prompt tokens, $0.50 / $2.50 above; cache $0.01; tokenizer ~30% more tokens; Haiku 4.5 was $1 / $5 / $0.10), GPT-6 Luna (steps up at 272K input tokens: 2x input, 1.5x output), GPT-6.1 Sol Ultrafast ($12 / $60 below 272K, $24 / $90 above; 6x Standard, 3x Fast), Mistral Large 4 ($0.68 / $2.09 on a 50% launch discount with no published end date; $1.36 / $4.18 list), OpenAI Decisions API ($0.10 input, $0 output, cache and reasoning)

Sources Anthropic pricing; OpenAI API changelog; Mistral AI

The decision-model category commoditizes before its largest entrant reaches GA

A category that did not have a name in August has, in six weeks, a typed interface (predicate, choice, score), four confirmed hosted vendors and one unverified (Perplexity's hosted price appears only in a third-party ladder), at least two downloadable implementations, a local-serving endpoint and a hosted price at or below four cents per million input tokens. The incumbent-then-clones sequence ran backwards: OpenAI's beta arrived late and priced highest. What is architecturally distinct is the output contract, not the network; every implementation whose base is disclosed is a classifier head post-trained on a shared 27B or smaller base (Microsoft's is a 9B Qwen, per its post), which is why the tree holds the category at the 3-of-6 test rather than cutting a branch. The infrastructure consequence is real regardless of placement: a verdict call emits a handful of tokens, needs little KV cache and no long reasoning trace, so volume that migrates here carries the lowest HBM and output-token intensity of any call type. If the approve/flag/block step of agent harnesses moves to this tier (Codex's Guardian-versus-Decisions telemetry is the first measurement of whether it can), aggregate request counts rise while tokens per request fall, which is the shape of demand that favours the accelerator vendor less and the network less than a chat-token boom would.

Examples

TypeSafe Jev (hosted, $0.042 since mid-September), OpenAI Decisions API on gpt-6-luna ($0.10 input, beta Oct 6), Microsoft-Decision-1 on Foundry ($0.042, Oct 9; post-trained Qwen3.5-9B per Microsoft's post, outside the graded set), Cloudflare Clef-flash ($0.038, reported), Perplexity pplx-decider-v1.1-27b (Apache 2.0, Decision Index 61.56), LM Studio 0.4.26 local /v1/decisions endpoint

Sources OpenAI API changelog; Microsoft Foundry; Perplexity; LM Studio

Open-weight label, closed artifact

Two Western frontier-scale announcements used the open-weight label for models that exist only behind an API, while four smaller or non-Western releases in the same week shipped repositories with licence metadata; Marc Pope (October 7) and TensorFeed (October 6) made the cannot-download point first. The tree's handling is the useful precedent for procurement: a model is placed on what can be fetched, with the licence read from the repository rather than the press release, and a promised release date is a watchlist item, not a property. Mistral's technical documentation for downstream providers names no open-weights licence; it describes the current on-prem terms as a bespoke and confidential self-deployment agreement and says the document will be updated if weights are released, which is why the AI Stack Weekly bets that the published licence will be neither Apache 2.0 nor MIT. Reflection's situation is thinner: no price, no card, vendor benchmarks revised three days after launch. For a buyer the test is binary: can your platform team run `huggingface-cli download` today. For Mistral Large 4 and Beam the answer is no, and the open-frontier scorecard slots stay with models where it is yes.

Examples

Mistral Large 4 (API preview Oct 6; weights promised for the end of October, licence not yet named), Reflection Beam (waitlisted API, no price; Apache 2.0 weights promised later in October; benchmark table revised Oct 8), Contrast: MiMo-V2.6-Pro, Kolibri-1, EmbeddingGemma 2, pplx-decider-v1.1-27b (downloadable)

Sources Mistral AI technical documentation for downstream providers; Reflection AI

Trillion-scale totals, sub-5% active fractions, four-week pretrains

Four MoE releases in two weeks cluster at an active fraction between 4% and 5% regardless of total size, from 78B to 1.05T. Four points are a cluster, not a settled design point, but they are consistent enough to plan serving around: per-token compute looks like a 25-50B dense model while the weights need the memory of the full total, which is why these models are marketed with 748 GB desktop boxes and single-node B200 or H200 footprints rather than with FLOPS. The training-time disclosures are the newer datapoint. Beam's 23.8T tokens in under four weeks on 6,144 GB300 NVL72 GPUs is the first frontier-scale pretrain disclosed on a Blackwell Ultra fleet (vendor-reported, report pending) and the only one of the two with a wall-clock attached; Mistral's from-scratch run on 3,800 Grace Blackwell GPUs in its own data centres is the week's other footprint disclosure, without a duration. The capital consequence is in the AI Stack Weekly: compute that produces a frontier model in four weeks is financed years ahead on somebody else's balance sheet, and the HBM that the 4-5% active fraction leaves idle is paid for either way.

Examples

Mistral Large 4 (1.05T / 52B active, 5.0%; 3,800 Grace Blackwell GPUs), Reflection Beam (501B / 23B active, 4.6%; 23.8T tokens in under four weeks on 6,144 GB300, vendor-reported), MiMo-V2.6-Pro (1.02T / 42B active, 4.1%; W40), Kolibri-1 (78.1B / 3.46B active, 4.4%; W40)

Sources Mistral AI; Reflection AI

Artificial Analysis Intelligence Index v4.3.2 (independent), cheap tier

Haiku 5.5 at max effort enters the sub-dollar tier five points above GPT-6 Luna; Mistral Large 4 enters level with Luna and behind both open Chinese models that cost less to run. Setting in parentheses is the reasoning-effort level; scores are not comparable across settings

Claude Haiku 5.5 (max)

43.4

GLM-5.3-Flash

42 (Capital and Compute's reading of the tracker)

DeepSeek V4.1 Flash

39 (Capital and Compute's reading of the tracker)

Mistral Large 4

38

GPT-6 Luna

38

Sources Artificial Analysis; GLM and DeepSeek rows via Capital and Compute

Terminal-Bench 4.0, vendor harness versus independent runs

Both new entrants score materially lower on independent runs than on their own harnesses; the gap is 6.4 points for Haiku 5.5 and 5.6 for Mistral Large 4. The mid-tier reference is Sonnet 5.5 at 70.6% on Anthropic's harness (W40)

Claude Haiku 5.5 (vendor, max)

39.2%

Claude Haiku 5.5 (Vals AI)

35.4%

Claude Haiku 5.5 (Artificial Analysis)

32.8%

Mistral Large 4 (vendor)

28.3%

Mistral Large 4 (Vals AI)

22.73%

Sources Anthropic; Mistral AI; Artificial Analysis; Vals AI

Artificial Analysis cost to run the Intelligence Index (independent), cheap tier

Artificial Analysis's Haiku 5.5 cost figure is provisional and excludes the tier step: about $0.21 per task at the sub-100K rate, three times Luna, as relayed by Capital and Compute from the tracker's Haiku 5.5 article, which also records Haiku at max effort using roughly 162K output tokens per task, about 3x Luna's; the row is carried as provisional rather than final and is outside the graded set. Mistral Large 4 costs about four times the two open Chinese models that outscore it

GPT-6 Luna (index 38)

$0.07 per task

Claude Haiku 5.5 (max; provisional, sub-100K rate, excludes tier step)

~$0.21 per task

GLM-5.3-Flash and DeepSeek V4.1 Flash (both above 38)

$0.25-0.27 per task

Mistral Large 4 (index 38, list price)

$1.13 per task

Sources Artificial Analysis (Haiku 5.5 article, artificialanalysis.ai/articles/claude-haiku-5-5) via Capital and Compute

DeepSWE v1.1, Reflection's own table (vendor-run, all rows)

Reflection's October 8 revised table places Beam below the three open Chinese models it is marketed against on both agentic-coding rows shared across all four (DeepSWE v1.1 tabulated here; Terminal Bench v2.1 is the other); the figures are Reflection's and have not been reproduced

DeepSeek V4.1 Flash

74.2%

Kimi K3

68.0%

GLM-5.3

61.0%

Reflection Beam

44.4%

Sources Reflection AI

Perplexity Decision Index (vendor-run) and MTEB Code (vendor-run)

Two vendor-reported specialist moves: Perplexity's v1.1 decider gains 5.2 points on its own index, and EmbeddingGemma 2 gains 9.9 points on MTEB (Code) over its predecessor

pplx-decider-v1.1-27b

61.56 (Decision Index)

pplx-decider-v1-27b

56.4 (Decision Index)

EmbeddingGemma 2 (740M)

78.68 (MTEB Code)

EmbeddingGemma 1

68.76 (MTEB Code)

Sources Perplexity; Google AI for Developers model card

2026-10-07 · Anthropic

Halves Sonnet 5.5 cache-read price to $0.10 per million (input, output and cache-write unchanged), citing a ~20% saving on most agentic tasks; adds monthly Claude Platform API credits of $100-200 for Max and up to $500 pooled for Team

For agent operators this is the week's only unambiguous cut, because cache reads are the meter long sessions hit hardest and the cut applies to the model most agent harnesses already default to. The API credits are a bundling signal: Anthropic is routing subscription revenue toward platform usage, which blurs the seat-versus-token boundary that The Application Layer tracks.

Sources Anthropic

2026-10-07 · Anthropic

Haiku 5.5 ships with breaking changes against Haiku 4.5: temperature, top_p, top_k, budget_tokens and prefill removed; the same text counts ~30% more tokens; retirement not before October 7, 2027

Migrations from Haiku 4.5 are not drop-in. Teams should budget a re-evaluation of any prompt that used sampling controls or prefill, re-measure token counts on their own corpus before accepting the 90% headline, and record the one-year retirement floor in vendor scorecards as a deployment guarantee.

Sources Claude Platform docs

2026-10-06 · OpenAI

Decisions API enters public beta on gpt-6-luna at $0.10 per million input tokens with no output, cache or reasoning charge, up to 10x faster than Luna via Responses (vendor claim); GA 'in the coming weeks'; API usage tiers collapse from five to three

Architects can now put a typed verdict call in a pipeline without running a classifier themselves, but the price is the category ceiling, not its floor, and the beta terms can change at GA. The tier simplification (Build $5, Launch $100, Grow $500) lowers the friction for small accounts and is the kind of change that precedes a volume push.

Sources OpenAI API changelog

2026-10-08 · OpenAI

GPT-6.1 Sol Ultrafast reaches GA for all API users at $12 / $0.60 cached / $60 per million below 272K input tokens ($24 / $90 above), with US and EU data residency; also in Codex and ChatGPT Work for Pro $500 and eligible Enterprise and Edu plans

Speed is now a purchasable SKU on the mid tier at a 6x premium. Operators should confine it to interactive paths where a human is waiting and keep batch and background agent work on Standard; the data-residency option makes it the first Ultrafast tier a regulated EU buyer can use.

Sources OpenAI release notes

2026-10-06 · Mistral AI

Mistral Large 4 preview priced at a 50% launch discount ($0.68 / $0.07 cached / $2.09) on a $1.36 / $0.14 / $4.18 list, with no published end date (a two-week duration circulates in secondary coverage of Mistral's changelog and is not in the graded set); weights promised for the end of October after red-teaming with cybersecurity partners and state authorities; no licence named

Budget the list price, not the sale. The red-teaming-with-state-authorities language and the bespoke self-deployment terms that govern on-prem access today are the sovereignty positioning made concrete, and they are the terms a European buyer should read before the weights land; the AI Stack Weekly bets that the published licence is neither Apache 2.0 nor MIT.

Sources Mistral AI

2026-10-08 · Local serving stack (vLLM, llama.cpp, LM Studio, Ollama)

vLLM v0.31.0 (717 commits) with breaking changes to multimodal kwargs and FP8 quantization flags; llama.cpp day-zero builds for the week's open releases; LM Studio 0.4.26 adds a local /v1/decisions-compatible endpoint; Ollama 0.40.2

Platform teams that pin vLLM should read the v0.31.0 migration notes before upgrading; the FP8 flag rename and the multimodal kwargs gate will break existing launch scripts. LM Studio's decisions endpoint means the four-cent verdict call can also be a zero-cent local one for on-device approval paths.

Sources GitHub (vllm-project/vllm releases)

Open (Mistral has published no end date)

Mistral Large 4 launch discount ends; price reverts to $1.36 / $4.18

The first test of whether the preview volume survives a 2x step; Mistral Studio usage and OpenRouter share are the public signals. A two-week duration circulates in secondary coverage and is not relied on here.

Oct 27-31

Mistral Large 4 weights and licence on Hugging Face; Reflection Beam weights, model card and technical report

Two deferred or provisional tree rows resolve on repositories: Mistral's openness field flips to open_weights if the files land, and the licence decides the AI Stack Weekly's bet that it is neither Apache 2.0 nor MIT (resolves November 15); Beam is placed or stays deferred on whether a card and checksums exist.

Q4 2026

Artificial Analysis publishes a tiered cost-per-task figure for Claude Haiku 5.5

The tracker's current figure is provisional and excludes the step above 100K tokens. When the tiered number lands, it settles whether Haiku's effective cost in agent loops is nearer Luna's $0.07 or the mid tier, which decides the sub-dollar scorecard row.

Coming weeks (per OpenAI)

Decisions API general availability and its price

A GA price at or below $0.05 is the AI Stack Weekly's bet (by January 31, 2027); a hold at $0.10 would mean OpenAI is selling the integration rather than the verdict, with the hosted tier already at or below four cents at TypeSafe, Microsoft and Cloudflare.

Q4 2026

Gemini 4 Argon developer availability outside the Fairwind Program

Carried from W40. The scorecard's closed-frontier challenger line still carries a model nobody outside the programme can call; the AI Stack Weekly's open prediction on Argon availability resolves on the Gemini API model list.

Open

Independent Terminal-Bench 4.0 and index runs for Reflection Beam; withdrawal count on OpenAI's 560 unverified mathematics manuscripts

Beam's vendor table was revised once already and no third party has run it; the OpenAI corpus's error rate, not its size, is what would make the unreleased model a tree entry when it ships.

• 2026-10-10: Issue 25 published. Three tree rows added (claude-haiku-5-5, mistral-large-4, embeddinggemma-2) and one updated (claude-sonnet-5-5 cache-read price). Reflection Beam, the decision-model fine-tunes (pplx-decider-v1.1-27b, Microsoft-Decision-1) and Liquid AI's reported d1 decision models are reviewed and deferred; GPT-6.1 Sol Ultrafast, the unreleased OpenAI mathematics model, Kandinsky 6.0 Video, Qwen-Image-2.1-Turbo and Sierra's in-house models are reviewed and excluded.
• 2026-10-10 (corrections, cycle 1): Haiku 5.5's sub-100K price is one tenth of Haiku 4.5's, not a quarter (a quarter is Anthropic's blended vendor estimate); Microsoft-Decision-1's disclosures replace an incorrect 'no base disclosed' claim; Jev is TypeSafe's model; Mistral Large 4 is removed from the sub-dollar framing and scorecard row (its output price exceeds $1 on sale and at list), its licence is described as not yet named with on-prem access under a bespoke self-deployment agreement, and the launch-discount duration is marked as unconfirmed; only Reflection's pretrain is a GB300 disclosure.
• 2026-10-10 (corrections, cycle 2): The Haiku 5.5 cyber-safeguards sentence now follows Anthropic's launch-page wording (more restrictive than Haiku 4.5's, somewhat less restrictive than its recent models'; biology safeguards matching Sonnet 5, Sonnet 5.5 and Opus 5); Artificial Analysis's Medium-setting figures for Haiku 5.5 are removed as unverified; the tokenizer effect is stated as 20-23% sooner (about 77K old-tokenizer tokens); the hosted decision-vendor count is four confirmed plus Perplexity unverified; the 'price war' frame is attributed to Dbggr and Latent Space; Mistral's active count is 52B only, its context is 1M per Mistral with Artificial Analysis measuring the preview at 524K; GLM-5.3-Flash (42) is added to the cheap-tier index table; the Mistral discount watchlist entry has no derived date. Figures used without a graded citation, labelled at point of use: Microsoft-Decision-1's disclosures (post-trained Qwen3.5-9B, 36 benchmarks, ~150,000 questions, six comparison models, ~35x P50; Microsoft's post); Artificial Analysis's provisional ~$0.21 Haiku cost per task and ~162K output tokens per task (~3x Luna), via Capital and Compute; the DeepSeek 39 and GLM 42 index scores (Capital and Compute); Jev and Microsoft-Decision-1 at $0.042 and Clef-flash at $0.038 (vendor pages via Willison and Dbggr); Willison's 1.25x tokenizer measurement; Anthropic's 75% blended estimate; the CTOL-reported $0.02 Perplexity hosted price (noted, not used); the wmedia.es 130K autocompact finding.
• Scorecard: a Sub-dollar hosted tier is added with Claude Haiku 5.5 as leader and GPT-6 Luna as challenger; the Edge / small leader moves from the 4B llama.cpp decision heads to EmbeddingGemma 2, the decision heads move to challenger, and Clef-flash drops off the row. The five other rows are unchanged.
• Every benchmark figure is labelled vendor-run or independent at the point of use; Reflection's table is presented as vendor-run on every row, and Artificial Analysis's Haiku 5.5 cost figure is carried as provisional (~$0.21, sub-100K rate, excludes the tier step) rather than as final.