Ornith-1.5: The Open-Weight Model Family That Writes Its Own Training Curriculum — and Claims to Beat Claude Opus 4.8

Ornith-1.5, the open-weight family from Ornith AI that claims to beat Claude Opus 4.8 on four benchmarks, shipped on August 19, 2026 — and the thing worth stopping over is the training loop, not the parameter count. The family is three models: Ornith-1.5-397B, a 397-billion-parameter mixture-of-experts flagship; Ornith-1.5-35B-A3B, a 35B MoE that activates 3 billion parameters per token; and Ornith-1.5-9B, a 9B dense model with a quantized mobile build. Every headline score here is vendor-reported: the numbers come from Ornith AI's own evaluation runs, averaged over five, and the one third-party tracker with a profile — BenchLM — labels all 18 of its benchmark rows provider-reported and keeps the family unranked until independent coverage exists. Here is what actually shipped, what "self-improving" really means, and what you can run today.

What shipped: three scales, MIT license, no API

Ornith AI — the group behind the open-weights Ornith-1.0 and associated with the DeepReinforce work on competitive-programming multi-agent systems (GrandCode) and GPU kernel optimization (CUDA-L2) — released the family on Hugging Face under an MIT license, with FP8, GGUF, MLX, and NVFP4 quantizations alongside the raw checkpoints. A 256K context window (262,144 tokens) applies across the family. There is no first-party hosted API and no token price; this is a self-host or local-runtime release, and the launch report says so plainly.

The three scales target three different realities:

• Ornith-1.5-397B — the "open-frontier bid": a 397B MoE aimed at the closed frontier on agentic and coding workloads.

• Ornith-1.5-35B-A3B — a 35B MoE activating only 3B parameters per token, for the middle ground: near-flagship capability with a fraction of a dense 35B's per-token inference cost.

• Ornith-1.5-9B — a dense 9B for local and on-device use, plus a quantized Ornith-1.5-9B-Mobile build the vendor says is ready to deploy on iPhone and Android.

The launch page above is the whole announcement: a technical report rather than a marketing post, which is consistent with the lab's profile.

The self-improvement claim, decoded

Ornith-1.0, released in June 2026, taught a model to generate the scaffolding around coding tasks — the harness, the decomposition, the orchestration. Ornith-1.5 extends that loop to task generation itself. In training, the model now does three things:

• Propose new, harder training tasks.

• Generate the task-specific scaffold used to solve and evaluate those tasks.

• Produce solution rollouts that become its next round of training data.

Reward flows through all three stages, jointly optimized with GRPO. Each proposed task is scored on validity (it must be executable and verifiable), frontier difficulty (the target solution-success rate is set at 0.2 — tasks the model usually fails but can still learn from), and novelty (diversity from everything already seen). The scaffold receives its own reward for alignment, reward fidelity, and resistance to reward hacking, and the pipeline includes anti-hacking safeguards: restrictions on touching the test environment, monitoring for restricted-path access, and a frozen adjudication model that overrides anomalous results.

The important caveat is in the word "self-improvement": it describes the training procedure, not the artifact. The downloaded model does not retrain itself on your laptop. What you get is a model whose training data was, unusually, generated by earlier versions of itself — which is exactly the kind of claim worth testing before it becomes dogma.

The benchmark claims, honestly labeled

All numbers in this section are Ornith AI's own, from the launch report, averaged over five runs, and not independently reproduced. The flagship's four claimed wins over Claude Opus 4.8:

• Terminal-Bench 2.1 (Terminus-2 harness): Ornith-1.5-397B 86.1 vs Claude Opus 4.8 85.0

• SWE-bench Verified: 86.0 vs 85.8

• WideSearch: 80.8 vs 72.9

• BrowseComp: 86.6 vs 84.3

Read those as vendor claims, not facts. The one flagship benchmark where Ornith AI does not claim a win is DeepSWE, 56.0 to Claude Opus 4.8's 59.0 — the report frames it as "on par," which is the honest way to read a loss. Other flagship figures from the same report: HLE 44.6 without tools and 56.1 with them, GPQA Diamond 92.8, and SWE-bench Pro 65.1.

The smaller two are strong on paper too: Ornith-1.5-35B-A3B reports SWE-bench Verified 79.0 and Terminal-Bench 2.1 (Claude Code harness) 68.5, while Ornith-1.5-9B reports SWE-bench Verified 70.6 and Terminal-Bench 2.1 47.0. Against other open-weights models in the same report, Ornith AI compares the 397B's 86.1 Terminal-Bench / 56.0 DeepSWE against DeepSeek-V4-Flash-0731's 82.7 / 54.4 and GLM-5.2's 81.0 / 46.2.

The one independent scoreboard — and why it is still provisional

BenchLM's profile for Ornith-1.5-397B is currently the only third-party page on the model. Its headline: a public score of 68.5 out of 100, ranked #20 of 221, with Agentic the strongest category at #13. But read the fine print, because it is doing real work — all 18 benchmark rows are marked provider-reported, and the profile states the model "does not yet have enough sourced coverage for a verified position." An unranked position on the verified list is not a verdict against the model; it is a statement that nobody has independently run it yet, which is exactly what you would expect for a days-old release.

That is the gap between the two numbers in this article: the vendor's 86.1 Terminal-Bench is a claim, and BenchLM's 68.5 is a score built entirely on vendor-reported rows. Neither is an independent measurement. The first lab to run Ornith-1.5-397B through a public harness — SWE-bench Verified on a controlled set, Terminal-Bench on a fixed harness version — will produce the first number that counts.

What you can actually run today

This is an open-weights release, so "run it" means pulling the weights. The Ollama catalog already lists ornith-1.5:9b (a ~6.6 GB build) and ornith-1.5:35b (a ~23 GB build), both at 256K context; the 9B build also carries image-input support in the catalog entry. The 397B is a different category of problem: a 397B-parameter MoE is not a laptop model, and any serious deployment means multiple GPUs and real orchestration.

The practical sweet spot for most teams is the Ornith-1.5-35B-A3B: 3B active parameters per token gives a MoE's capability at a fraction of a dense 35B's per-token cost, and the 23 GB Ollama build runs on a single high-VRAM card or a small cluster. The 9B is the one you put on a phone.

That "3B active" property is also where the routing economics get interesting. Active-parameter MoEs price far below their total size, and when providers adopt a model like this the per-token price tends to fall fast — which is the case for buying through a router that passes provider list prices through at 0% markup, rather than a single vendor you negotiate with once. OrcaRouter does exactly that across 200+ models, so a provider price cut on a new open-weights model is live on our side the same day. And for a model that shipped with nothing but vendor benchmarks, the failover argument matters most: a routing layer lets you point a test path at a brand-new model and fall back to a proven one the moment it stalls — trying an unproven release without betting a production path on it.

What is still missing

• No hosted API. If you want Ornith-1.5 as a service, it does not exist yet; every option is self-host or local.

• No independent evaluation. BenchLM keeps the family unranked; no lab has published a controlled run.

• No pricing to reason about. There is no token rate, so the "cheap open frontier" case is entirely about your own GPU economics.

• The harnesses are mixed. Terminal-Bench has multiple variants, and the report's own numbers span them: the 397B's 86.1 is on the Terminus-2 harness, the 35B's 68.5 on the Claude Code harness. Different harnesses are different games, which makes apples-to-apples comparison harder than the headline table suggests.

What to watch next

The durable story is not "open model beats Claude Opus 4.8" — that is an unaudited claim a day old. It is that a lab has closed the loop on self-generated training data and published the weights for scrutiny, and that is the part worth watching. The things that would turn this from a promising release into a trusted one: a first independent run of the 397B on SWE-bench Verified and a fixed Terminal-Bench harness; the evaluation code actually shipping; and a hosted route appearing so the 35B-A3B's cost profile can be measured against its claims. If the numbers hold, Ornith-1.5-35B-A3B is the open-weights price/performance story of the quarter. If they do not, the honest part — the self-improvement method and the MIT weights — survives either way.