brianletort.ai
All issues

The Model Pulse

Issue 18 · Week 34 of 2026.

/Weekly read/~6 min read/Public sources onlyDownload brief

The Big Read

OpenAI cut Sol by more than a fifth for three months while an unknown lab shipped an MIT frontier — the same week the industry conceded that vendor-reported is what the top of the board looks like now

The thesis this issue defends

Two things happened in the software layer this week that only make sense read together. On Thursday Ornith AI, a lab nobody had heard of a fortnight ago, published Ornith-1.5 under the MIT license — a 397B mixture-of-experts flagship with 35B active parameters, a 35B dense sibling, and a 9B distilled variant, benchmarked on the lab's own harness at 89.7% on Terminal-Bench 2.1 and 61.4% on DeepSWE. The next day OpenAI announced a token price cut of more than 20% on GPT-5.6 Sol, held for three months. One vendor said the top of the coding leaderboard is now open, released under a permissive license, and available to anyone. The other vendor said the current price of frontier reasoning tokens is negotiable and it is willing to say so out loud, for a fixed period, in writing.

The honest reading is that neither the leaderboard claim nor the price claim is settled. Ornith AI's Terminal-Bench figure comes from the lab's own harness, which has not been audited by any third party, and no independent tracker — Artificial Analysis, Vals, LMSYS — had ranked Ornith-1.5 by publish time. This publication is adding the model to the tree because the artifact is real and downloadable, and it is recording it at medium placement confidence with a notable field saying so plainly. If Artificial Analysis or Vals reproduces the score inside the next four weeks, the classification is trivial to lift. If they do not, that itself is the finding, and it belongs in a procurement document rather than a headline.

OpenAI's move is the more consequential one for buyers this quarter. A promotional 20%-plus cut on the flagship reasoning model, published as a dated three-month change, is the same behaviour Google put on the Gemini 3.7 Flash pricing page last week: cheap inference sold with an expiry. This publication's precision rule is that a price with a published end date is a promotional variable, not the durable rate — the durable rate is whatever the vendor's willing to write on a contract that spans the reversal. Every agent business case whose payback crosses November should model the post-promo price, and every vendor cost comparison built on this week's per-token headline should carry the date attached.

A third release rounds out the layer without changing it. DeepSeek published DeepSeek-V4-Flash-Vision-Exp on Friday, an experimental API-only vision-capable variant of its Flash tier with no weights and no full-quality Pro comparator, aimed at multimodal harness developers testing the API surface. This publication is not adding it to the tree — models.yaml does not have a pattern for experimental API-only rows, and creating one for a variant with no weights would set a precedent the taxonomy would then have to defend against every future API-only preview. The Pulse notes it and moves on.

The absent number is the Artificial Analysis Intelligence Index. No new AA Index release landed in-window, so Grok 4.6, GPT-5.6 Sol Max and Claude Opus 5 are still separated by two points of a measurement last refreshed on August 6. Coverage will keep quoting positions on the board this week; none of them changed. And CyberGym remains where GLM-5.3 and DeepSeek-V4 Pro 0813 left it last week — vendor-reported, unreproduced, and cited interchangeably in trade press.

One finding on the served-versus-downloadable split from last week worth pulling forward: Z.ai has still not published GLM-5.3 weights to Hugging Face. The two-week window announced with the model closes at the end of this issue's coverage period. If nothing lands in the next seven days, the benchmark-first-weights-later pattern this publication flagged in W33 gets one more instance and moves toward being a durable vendor behaviour rather than a one-off. The tree keeps GLM-5.3 recorded as gated until weights are confirmed.

Tree delta

What changed in the tree.

1 model added, 0 updated.

1 model row added and 0 updated. Ornith-1.5-397B enters as a medium-confidence, open-weights MoE entry with every headline benchmark labelled vendor-reported and awaiting independent reproduction. DeepSeek-V4-Flash-Vision-Exp is deliberately kept out of the tree because models.yaml has no pattern for experimental API-only rows and one variant is a poor precedent to set.

Added (1)

  • ornith-1-5-397b

Updated

None this period.

The Ornith-1.5 entry carries the smallest evidence footprint the tree has admitted for a frontier claim this year: one aggregated secondary source and vendor-run harness scores. It is recorded because the weights are downloadable and MIT-licensed, which makes the artifact real regardless of the score's provenance. Reclassify to placement_confidence: high once an independent tracker reproduces at least one headline number. If the two-week Z.ai window on GLM-5.3 weights closes without a Hugging Face release, next issue will discuss whether to move GLM-5.3 off gated on the grounds that a promised-but-unshipped weight is a different classification than a shipped one.

Explore the LLM Evolutionary Tree

Frontier movements

Flagship-class releases.

2 releases this period.

Vendor-stated frontier capability. The releases that reset the closed-source ceiling.

  • /OpenAI/Frontier/Reasoning

    GPT-5.6 Sol (August price recut)

    Token price on the flagship Sol tier cut by more than 20% for three months, disclosed as a dated promotional window rather than a durable rate change

    The most economically consequential model event of the week and the one most likely to be misread as a permanent cut. OpenAI published the end date and the reversal at the same time as the reduction, which follows the pattern this publication logged with Gemini 3.7 Flash: cheap inference sold with a published expiry. Buyers should model any agent business case whose payback crosses the November reversal against the post-promo rate rather than the promotional one, and finance leaders should note that a three-month window is short enough to sit inside a single quarterly board cycle — which is likely the point. The capability did not move; only the meter did.

    OpenAI GPT-5.6 announcement page

  • /DeepSeek AI/Specialist/Multimodal

    DeepSeek-V4-Flash-Vision-Exp

    API-only experimental vision-capable variant of the Flash tier — no weights, no full Pro comparator, positioned at multimodal harness developers

    Recorded on the Flash line rather than added as a new frontier row because models.yaml has no pattern for API-only experimental variants and creating one for a preview would set a precedent that fills the tree with previews. The interesting signal is what it signals about DeepSeek's release calendar: a vision experiment shipped as API-only inside two weeks of the silent V4 Pro 0813 endpoint swap suggests the vendor is iterating on the served surface faster than on the weights it publishes. Teams building multimodal harnesses should test but avoid pinning production behaviour to the experimental endpoint.

    DeepSeek API documentation

Open weights

Open-frontier and open-source drops.

1 release this period.

Open-weights releases that change procurement options. Pull these into pilot when score parity meets license parity.

  • /Ornith AI/Open frontier/MoE

    Ornith-1.5

    397B/35B active MoE flagship (plus 35B dense and 9B distilled siblings) under MIT license, with vendor-run harness scores of 89.7% on Terminal-Bench 2.1 and 61.4% on DeepSWE

    The evidence weight on this one is unusually thin and the license unusually clean, which is exactly the combination that makes it hard to file. MIT with no user gate, no naming rule and no acceptable use policy is materially more permissive than the Llama license and slightly more permissive than Apache 2.0 in effect. Every headline number is vendor-reported on a vendor-authored harness, and no independent tracker had listed Ornith-1.5 by publish time — so buyers should treat the score with the same care as any vendor benchmark and treat the artifact as real for licensing decisions but unproven for capability decisions. The 9B distilled sibling is the more immediately useful piece for teams evaluating local-first agentic tooling.

    RuntimeWire aggregation of Ornith AI announcement and repository

Architecture watch

Patterns to track.

3 patterns reshaping the canopy.

Architectural patterns that crossed multiple vendors this period. Each pattern lists exemplar releases and what it changes for deployment, cost, or capability.

  • Frontier price cuts are being sold with a written expiry date attached

    GPT-5.6 Sol -20% for 3 monthsGemini 3.7 Flash Jan 1 doublingDeepSeek peak/off-peak

    OpenAI made the pattern explicit this week: a headline cut of more than 20% on the flagship reasoning model, held for three months, with the reversal date published at the moment of the cut. That is the third dated promotional inference rate a major lab has shipped in five weeks. The read is that top-tier inference pricing has moved from list-price competition to promotional periods that fit inside a customer's quarterly planning window — which is easier to close deals against and much harder to build a durable margin on. Enterprise buyers should be writing model contracts with rate protection through the next reversal date and asking every vendor for one, because promo windows have become the durable form of the disclosure.

    OpenAI GPT-5.6 announcement, Google Cloud pricing history

  • The evidence bar the tree is being asked to admit at got noticeably lower

    Ornith-1.5 vendor Terminal-BenchGLM-5.3 promised weightsDeepSeek 0813 silent swap

    Three of the last two weeks' most-cited coding scores — Ornith-1.5, GLM-5.3, DeepSeek-V4 Pro 0813 — share the same evidence structure: the vendor ran the harness, the score sits at or near the top of the board, and no independent lab has reproduced it. That is not fraud, and none of these vendors dispute the labelling. It is a structural change in how a top-of-board claim now enters public awareness, and it means a procurement document quoting the ranking without the version, the harness and the vendor-reported label is measuring hype rather than capability. Treat every leaderboard row this quarter as provisional until an independent tracker reproduces the score, and treat 'the leader on X' phrasing as unverified where the leader's harness is the one measuring it.

    Ornith AI, Z.ai, DeepSeek observed release pattern

  • API-only experimental tiers as a shipping-surface for multimodal work

    DeepSeek-V4-Flash-Vision-ExpGemini prior experimental Flash previewsOpenAI Ultrafast preview

    DeepSeek followed Google's pattern of shipping vision or reasoning previews as an API-only tier separate from the weights track. This publication is choosing not to add these entries to the taxonomy as new frontier rows because the pattern is not durable inside a model card — an experimental endpoint can be retracted or renamed without a release note, and models.yaml would end up with a fossil layer of preview names. What the pattern does signal is that the served surface has become where the vendor iterates on modality, while the weight release track is where architecture is disclosed. Buyers building multimodal harnesses should assume the API surface will move faster than any published weight and design their evals to accommodate that.

    DeepSeek API documentation, Google Cloud Gemini experimental history

Benchmark moves

Where the leaderboard moved.

4 benchmarks shifted.

Benchmark deltas that change a procurement read. Scores reflect public leaderboards or vendor model cards as of publication.

  • Artificial Analysis Intelligence Index

    No new Index release in-window. The board is unchanged from the August 6 v4.1.1 refresh, and any coverage quoting new Index positions this week is quoting the prior refresh under a fresh headline.

    • Claude Opus 563 (as of Aug 6)
    • Grok 4.661 (as of Aug 6)
    • GPT-5.6 Sol Max61 (as of Aug 6)

    Artificial Analysis Index v4.1.1 release history

  • Terminal-Bench 2.1 (vendor-reported rows)

    Ornith AI publishes 89.7% on its own harness, entering the top of the board on unaudited numbers. The comparability question is whether the vendor's harness measures the same thing the Anthropic-authored public evaluator measures, and this publication is not yet convinced it does.

    • Claude Fable 5 (v2.1)88.0
    • DeepSeek-V4 Pro 0813 (v2.1, vendor)87.9
    • Ornith-1.5-397B (vendor harness)89.7
    • GLM-5.3 (v3.0, vendor)top open-source claim (figure undisclosed)

    Ornith AI harness, Z.ai release notes, DeepSeek OpenLLMStack aggregation

  • DeepSWE (software engineering, vendor rows)

    Ornith-1.5 publishes 61.4% on the vendor harness, slotting between the closed-frontier ceiling and the prior open-source line. No independent reproduction attached.

    • GPT-5.6 Sol Max73%
    • Ornith-1.5-397B (vendor)61.4%
    • DeepSeek-V4 Pro 0813 (vendor)62.7
    • Grok 4.6 (post-training)65.9%

    Ornith AI release notes, prior Pulse rows

  • LMSYS Arena, Text listings

    Grok 4.6 continues to hold third position from last week; no new frontier listings landed in-window. Ornith-1.5 has not appeared on Arena, which is consistent with its lack of an independent tracker read.

    • Claude Opus 5Top
    • GPT-5.6 Sol MaxSecond
    • Grok 4.6Third

    LMSYS Chatbot Arena leaderboard, snapshot at publish

Tier scorecard

Who leads, who pushes.

6 tiers · leaders as of Aug 22, 2026.

A snapshot of leader-vs-challenger by tier. Useful for procurement shortlists when matching workload to model class. Pair with the benchmark moves above for the underlying scores.

  • Closed frontier

    Leader: Claude Opus 5

    Challenger: Grok 4.6

    Unchanged on capability from W33's read: Opus 5 holds the top of a board that did not refresh this week. The material change is on price, not on position — Sol's three-month promotional cut narrows the cost gap that separated Grok 4.6 from Sol Max on comparable tasks. Decide the tier on cost per completed task inside the promo window, and re-decide it in November.

  • Open frontier

    Leader: DeepSeek-V4 Pro 0813

    Challenger: Ornith-1.5-397B

    The challenger enters on license terms, not on independently verified capability. Ornith AI ships under MIT with 35B active MoE parameters and a 9B distilled sibling; DeepSeek's served build is still not identical to its downloadable weights. The pick this quarter depends on whether the buyer weights license permissiveness (Ornith) or served-endpoint reliability with a paid API (DeepSeek).

  • Reasoning

    Leader: Claude Opus 5

    Challenger: GPT-5.6 Sol Max

    No capability change this week. The efficiency read from W33 stands — Opus 5 remains materially more expensive per completed long-horizon task than Grok 4.6 or Sol Max — and Sol's promotional cut applies to token price, not to turn efficiency, so the durable efficiency ranking is unchanged.

  • Coding

    Leader: Claude Fable 5 Max

    Challenger: GPT-5.6 Sol Max

    Vendor-reported Terminal-Bench 2.1 numbers from Ornith and DeepSeek are not directly comparable to the audited Fable and Sol numbers because the harnesses differ. Fable 5 Max still leads CursorBench v3.2 and FrontierCode v1.1 Extended; Sol Max still leads DeepSWE v1.1. Ornith and DeepSeek's own results are on the board but not scored against them here until an independent lab reproduces one.

  • Multimodal

    Leader: Gemini 3.7 Flash

    Challenger: Muse Spark 1.2

    Unchanged. DeepSeek's V4-Flash-Vision-Exp is API-only and experimental, so it does not enter the row. Gemini 3.7 Flash's January 1 doubling still applies and this publication is still counting the row on introductory pricing.

  • Edge / small

    Leader: Muse Glimmer

    Challenger: Ornith-1.5-9B (distilled)

    Ornith AI's 9B distilled variant is the first potentially frontier-adjacent 9B open-weight release under MIT this quarter. It sits below Glimmer on Intelligence Index reads (once Glimmer is reproduced independently on the same harness) but likely above on Terminal-style tasks per vendor scores. Reclassify when a third-party benchmark lands.

Vendor signals

Pricing, gating, deprecation.

5 non-release signals worth tracking.

The non-release moves that shift vendor risk — pricing, deprecations, gating decisions, license changes — with a one-line procurement read.

  • /OpenAI

    Announced a token price reduction of more than 20% on the GPT-5.6 Sol tier, effective for three months from the announcement date, with the reversal date and return-to-list rate published at launch

    This is the first frontier US lab to explicitly time-box a price cut on its flagship reasoning model, and the shape matches the Gemini and DeepSeek pattern this publication has been tracking. Enterprises negotiating enterprise agreements this quarter should ask for rate protection through and beyond the November reversal, because the cut is real for the promo period and the return-to-list is equally real after.

    OpenAI GPT-5.6 announcement page

  • /OpenAI

    Announced ChatGPT ads to launch in 31 EU countries with a formal rollout date of August 24, plus a separate ChatGPT for Teens tier, positioning consumer surfaces as a distinct product line from the enterprise API

    Two consumer moves that add up to a distribution story rather than a capability story. Advertising on the free tier is the first material step toward funding consumer inference through a channel other than subscription revenue, and the Teens tier is a compliance and safety product carved out ahead of the more likely EU AI Act enforcement wave. Enterprise buyers should treat neither as relevant to their contracts and note that OpenAI's product surface has visibly split into consumer and enterprise SKUs faster than any competitor's has.

    OpenAI blog and press coverage

  • /Ornith AI

    First public model release: Ornith-1.5 family under MIT license across a 397B/35B active MoE flagship, a 35B dense sibling, and a 9B distilled variant, published with vendor-run harness scores at or near the top of the board

    The evidence structure is the tell rather than the score. A new lab shipping an MIT-licensed frontier claim inside a week of its first observed public footprint sets a very specific procurement problem: the license is clean enough to matter for legal review, and the capability claim is thin enough that no serious infrastructure decision should ride on it until an independent tracker reproduces at least one headline number. Treat the weights as usable for evaluation and the scores as unproven.

    RuntimeWire aggregation of the Ornith AI release

  • /DeepSeek AI

    Published DeepSeek-V4-Flash-Vision-Exp as an API-only experimental tier with no accompanying weights and no full Pro-tier vision comparator

    The Pulse is reading this as a shipping-surface signal rather than a capability signal: DeepSeek is iterating multimodal work on the API surface rather than on the weight track, which is consistent with the silent 0813 endpoint swap two weeks ago. Multimodal harness developers may test it, but no production system should depend on an experimental endpoint that ships without an announcement or a stability commitment. Recorded but not added to the tree.

    DeepSeek API documentation

  • /Z.ai

    GLM-5.3 weights still not on Hugging Face at publication; Z.ai's own two-week promise window from the August 14 announcement closes at the end of this issue's coverage period

    This publication does not move a model out of gated on promises. If nothing lands in the next seven days, next issue will discuss whether the benchmark-first-weights-later pattern rises to the level of a durable vendor behaviour worth naming, and whether models.yaml should distinguish 'weights promised' from 'weights shipped' as separate rather than transitional states.

    Z.ai Hugging Face repository status at publish, W33 Pulse

Watchlist

On the radar next.

6 catalysts to watch, starting Aug 26.

Specific model-side catalysts in the next 7–30 days that would change the read materially. Watching these tells us whether the canopy is widening or thinning.

  • Aug 26

    NVIDIA Q2 FY2027 earnings, data-center revenue, and any Rubin Ultra memory statement

    The single largest scheduled disclosure in the quarter, and the earliest chance for NVIDIA to confirm or deny the 8-stack HBM4E de-spec reporting that has been carried unresolved for two weeks. Any official memory spec commentary would resolve W33's second-noise call directly, and would be the first primary source on Rubin Ultra configuration rather than a rumor-based one.

  • Aug 22 - Sep 5

    Independent reproduction of Ornith-1.5 headline scores on Artificial Analysis or Vals

    The bar this publication is applying is that an MIT-licensed frontier claim from an unknown lab needs at least one independent leaderboard read before the tree lifts placement confidence. Two weeks is a reasonable window for a public tracker to test a downloadable model, and either a confirmation or a contradicting result is decision-relevant for procurement.

  • Aug 29

    Z.ai two-week window on GLM-5.3 weights closes

    The announced window ends on or around this date. Whether the weights land is the difference between a shipped open-source model and an announcement, and this publication will name the pattern as durable if a second frontier open-source release ships weights on a similar delay this quarter.

  • Nov 21

    OpenAI GPT-5.6 Sol promotional pricing window closes

    The dated reversal on the 20%-plus cut. Watch whether OpenAI extends, replaces or reverts on the exact date — an extension would soften the durable-rate reading of the cut, a reversion would confirm it. Any enterprise contract signed against the promotional rate needs a plan for that Monday.

  • Sep-Oct

    Artificial Analysis Intelligence Index v4.2 or v4.1.2 refresh

    The Index has not refreshed since v4.1.1 on August 6, which means all Index positions currently in trade press are two-plus weeks stale. The next refresh will be the first opportunity to see whether Ornith-1.5, GLM-5.3 or the DeepSeek 0813 build enter the ranked board on independent scoring.

  • Sep-Oct

    Anthropic in-house silicon commentary and shipping timeline signal

    Anthropic hired Amir Salek (a TPU program veteran) this week for in-house silicon per secondary reporting; the Pulse does not carry the hire as a model event but flags any subsequent official confirmation or Trainium-independence commentary from Anthropic as the next relevant data point. A durable capability shift shows up as vendor guidance to customers, not as a personnel move.

Edits this issue

  • Tree-delta admission bar tightened: the Ornith-1.5 entry is recorded at medium placement confidence with a notable field explicitly citing the vendor-run harness and absent independent reproduction, in line with the rule that low- and medium-confidence entries carry their reason in the row rather than in the changelog.
  • Experimental API-only variants are no longer eligible for treeDelta.added on their own — DeepSeek-V4-Flash-Vision-Exp is noted in vendor signals but not added to the tree, on the grounds that a preview endpoint should not consume a stable id in models.yaml until it ships weights or graduates.
  • Promotional inference pricing is now treated as a shipping-surface event rather than a rate change: the OpenAI Sol cut is logged with its expiry date attached, and the durable rate is defined as whatever survives the reversal, consistent with the precision rule applied to Gemini and DeepSeek in prior issues.

About The Model Pulse

A weekly read on the software side of the AI stack. Anchored to the LLM Evolutionary Tree, which the brief annotates each week. The cross-stack flywheel (capital, hardware, networking) is covered in The AI Stack Weekly.

Authorship and sources

Compiled from public model cards, vendor blogs, leaderboards, and official lab announcements. Written by Brian Letort. Independent analysis. Not investment guidance.

Operate. Publish. Teach.