Frontier Models — the 2026 Landscape

Honest snapshot as of 2026-06-16. Numbers below are cross-checked across at least two sources where possible; single-source or self-reported figures are flagged inline. Benchmark scores in 2026 are noisy — vendors report different harnesses, effort levels, and scaffolds for the "same" benchmark, so treat any single number as ±a few points and a relative signal, not a verdict. Where a figure could not be verified against a primary source, this article says so rather than inventing one.

What it is

The "2026 frontier-model wave" is the Q2-2026 cluster of releases that pushed the capability ceiling and — more importantly for builders — collapsed the price of frontier-grade coding and reasoning. Three things happened almost simultaneously:

1. The proprietary leaders sharpened. Anthropic shipped Claude Opus 4.8 (2026-05-28), reclaiming the #1 spot on the Artificial Analysis Intelligence Index, then released Claude Fable 5 (2026-06-09) — its first publicly-available Mythos-class model. OpenAI shipped GPT-5.5 (2026-04-23, codename "Spud") with a 1M-token API context and a Codex integration that now drives a majority of its own engineering.
2. A Chinese open-weight wave landed in a three-week window (≈Apr 7–24): DeepSeek V4 (1.6T MoE, MIT), Kimi K2.6, Qwen 3.6, GLM-5.1, MiniMax M2.7, plus Gemma 4 from Google — collectively the first time open weights credibly contested the frontier on agentic coding, at 15–30× lower token cost.
3. A genuine science landmark: an internal OpenAI reasoning model autonomously produced a counterexample to an 80-year-old Erdős conjecture (announced 2026-05-20), verified by a companion paper with external mathematicians — the first time a prominent open problem central to a subfield was settled autonomously by AI (with heavy caveats, below).

Why it matters

For OpenAlice this is not trivia — it directly reshapes the provider/model choice, [[model-routing]] policy, and the cost-ladder that Alice and the lab rigs run on. The headline of 2026 is not "the smartest model got smarter" (it did, marginally). It is that open-weight models now sit one rung below the proprietary frontier on coding while costing an order of magnitude less, and that the proprietary leaders bifurcated into a public tier (Opus 4.8 / GPT-5.5) and a gated, safety-classified super-tier (Fable 5 / Mythos 5). A router that treats "frontier" as one bucket is now leaving both money and capability on the table.

Capabilities & positioning

Anthropic — Claude Opus 4.8 + the 4.X family + Fable 5

• Opus 4.8 (2026-05-28): #1 on the Artificial Analysis Intelligence Index at 61.4 (+4.1 over Opus 4.7, ≈+1.2 ahead of GPT-5.5 xhigh — verified against the Artificial Analysis Opus 4.8 article). Reported 88.6% SWE-bench Verified and 69.2% SWE-bench Pro (up from 64.3% for Opus 4.7). Pricing unchanged at $5 / $25 per M input/output tokens, 1M context, with a new "3× cheaper fast mode." Anthropic frames it as a "modest but tangible" step over 4.7, with the notable claim that it is ~4× less likely to let a code flaw pass unflagged — a reliability/honesty gain rather than a raw-capability leap. - Caveat: the 88.6% SWE-bench Verified figure is widely repeated in secondary roundups (Vellum, codersera, computingforgeeks) but the Artificial Analysis writeup itself headlines the Index score, not SWE-bench. Treat 88.6% as vendor/secondary-grade, not independently re-run here.
• The 4.X family (4.6 → 4.7 → 4.8) is the public daily-driver tier: same price band, steady per-point gains on coding and agentic benchmarks, 1M context.
• Claude Fable 5 (2026-06-09) + Claude Mythos 5: Fable 5 is the first publicly-available Mythos-class model — Anthropic calls it state-of-the-art "on nearly all tested benchmarks," claiming >10% over Opus 4.8 on some benchmarks and being "first to break 90%" on its core analytics eval. On the Artificial Analysis Index a Fable 5 (Max Effort, Opus-4.8 fallback) configuration topped the board at ~65. Pricing $10 / $50 per M (from June 23). - Two honest caveats: (1) Anthropic published qualitative comparisons — the page does not disclose exact SWE-bench/GPQA numbers, so the ">10%" and "90%" claims are self-reported and benchmark-unspecified. (2) Within ~3 days of release a US government export directive temporarily forced Fable 5 offline — availability is politically contingent, and Mythos 5 ships without the safety classifiers Fable carries. See [[ai-safety-and-jailbreaks]].

OpenAI — GPT-5.5 + Codex + the Erdős landmark

• GPT-5.5 (2026-04-23, "Spud"): first OpenAI model with a 1M-token API context. Reported 82.7% Terminal-Bench 2.0 (state-of-the-art on terminal/CLI coding), 58.6% SWE-bench Pro, 84.9% GDPval, while using fewer tokens than GPT-5.4. On the Artificial Analysis Index it scores ~60.2 — a close #2/#3 behind Opus 4.8. Positioning: leads on terminal/agentic CLI coding, trails Opus 4.8 on SWE-bench Pro and the overall Index. - Caveat: note the benchmark split — "best at coding" depends entirely on which coding benchmark. GPT-5.5 wins Terminal-Bench; Opus 4.8 wins SWE-bench Pro; DeepSeek V4 leads on SWE-bench Verified among open weights. There is no single "coding king."
• Codex (the GPT-5.x-Codex line, latest public 5.3-Codex): now used weekly by >85% of OpenAI staff; the substrate behind much of OpenAlice's own delegation flow (Codex OAuth is our default provider). Relevant to [[agentic-loops]] and [[tool-use-function-calling]].
• Erdős landmark (2026-05-20): an internal general-purpose reasoning model produced a counterexample to a 1946 Erdős conjecture in discrete geometry, with a companion verification paper co-authored by external mathematicians (including Thomas Bloom, who debunked OpenAI's failed Oct-2025 claim). Princeton's Noga Alon called it "an outstanding achievement." - Honest framing: this is a real autonomous result, distinct from the discredited Oct-2025 "solved 10 Erdős problems" episode (which was literature-retrieval, not proof). But it is one problem, from an internal (unreleased) model, and "frontier math" success does not generalize to reliable everyday reasoning. The arXiv writeup (2601.07421) covers a related Erdős #728 Lean proof — readers should not conflate the several distinct Erdős-problem threads circulating in 2026.

DeepSeek V4 — the open-weight coding spearhead

• Release 2026-04-22, full MIT weights on HuggingFace. Two checkpoints: V4-Pro (≈1.6T total / ~49B active, 1M context) and V4-Flash (≈284B / ~13B active). Architecture continues DeepSeek's sparse-MoE lineage with a hybrid attention design (Compressed Sparse Attention + a new heavily-compressed long-context head). See [[deepseek-architecture]] and [[mixture-of-experts]].
• Coding: V4-Pro-Max ≈ 80.6% SWE-bench Verified — the highest open-weights entry, reportedly tied with Gemini 3.1 Pro — plus ~93.5% LiveCodeBench and ~90.1% GPQA Diamond (vendor-reported). Output around $0.87/M — roughly 1–2 orders of magnitude cheaper than Opus 4.8 / Fable 5. - Caveat: the 80.6% is DeepSeek-reported on the "Pro-Max" reasoning configuration; independent re-runs typically land a few points lower, and "tied with Gemini 3.1 Pro" comes from secondary roundups, not a head-to-head we verified.

The Chinese open-weight wave (Kimi K2.6, Qwen 3.6, GLM-5.1, MiniMax M2.7, Gemma 4)

A dense April cluster — these are the models that made "open weights" a serious router option:

• Kimi K2.6 (Moonshot, 2026-04-20): reportedly the first open-weight model to beat GPT-5.4 (xhigh) on SWE-bench Pro; ~66.7% Terminal-Bench 2.0; notable stability — a published 13-hour / 4,000+ tool-call session, a long-horizon ceiling other open models don't reach (see [[agentic-rl-long-horizon-coding]]).
• GLM-5.1 (Z.ai): strong agentic web-dev; independently-reported Code Arena Elo ~1,530 (~3rd globally on agentic web dev), reflecting real developer-preference head-to-heads.
• Qwen 3.6 (Alibaba): 1M-token context (Qwen 3.6 Plus), plus small MoE variants (e.g. Qwen3.6-35B-A3B) good for self-hosting and [[small-language-models]] use.
• MiniMax M2.7: agentic self-improvement angle — reportedly ran 100+ internal rounds optimizing its own scaffold.
• Gemma 4 (Google): the Western open-weight entry rounding out the field for on-device / small deployments.
• Cross-cutting caveat: the "15–30× cheaper than international peers" claim is real in direction but quoted from a single roundup (TokenMix); exact multiples vary by workload, host, and whether you self-host or rent. The benchmark wins above are largely vendor- or arena-reported and subject to [[benchmark-contamination]] concerns — open-weight labs train on public eval-adjacent data, and arena Elos reflect preference, not correctness.

Comparison table

Index = Artificial Analysis Intelligence Index (composite of 10 evals; higher = better). Coding column cites the strongest verified-ish coding number per model — benchmarks differ between rows (SWE-bench Verified vs Pro vs Terminal-Bench), so columns are not directly comparable. Treat as positioning, not a ranking. Prices are USD per M input/output tokens.

¹ Secondary-source/vendor figure, not re-run here. ² Exact GPT-5.5 / Gemini per-token prices not re-verified in this article — check the live provider page before routing on cost. ³ DeepSeek V4 was not listed on the AA Index in the snapshots checked; absence ≠ low score.

Honest caveats & unknowns

• Benchmark soup. "SWE-bench" alone now means Verified, Pro, Lite, and various harness/effort configs. Vendors headline whichever they win. Do not compare a "SWE-bench" number across two vendors without confirming the exact variant and scaffold. See [[llm-evaluation]] and [[benchmark-contamination]].
• Self-reported vs independent. Fable 5's ">10% over Opus 4.8" and "90% analytics" are Anthropic-stated with no published per-benchmark numbers. Open-weight coding wins are largely vendor/arena-reported. Independent re-runs (e.g. via the lab's own bench.blal.pro↗) are the only numbers we should fully trust.
• Availability is political. Fable 5 went offline within 3 days under a US export directive; Mythos 5 is gated. Open-weight Chinese models face the inverse risk (Western procurement / policy). A router betting on a single proprietary super-model is exposed to non-technical outages.
• Math landmark ≠ general reliability. The Erdős result is real and externally verified, but it is one problem from an unreleased internal model. It says little about day-to-day agentic correctness.
• Prices move weekly. Every per-token figure here decays fast. Re-check the provider page before encoding a cost threshold.
• "Cheap" open weights aren't free to run. $0.87/M assumes a host. Self-hosting a 1.6T MoE has real infra cost; the 15–30× advantage is workload-dependent.

How it connects to OpenAlice

OpenAlice's defaults (per lab memory, 2026-06-16): provider = Codex OAuth, default model gpt-5.5 on mvp (prod still gpt-5.4 pending next deploy), with E2E standard Codex OAuth gpt-5.4. The 2026 landscape gives the lab three concrete levers:

1. Cost-ladder routing. The open-weight wave is precisely what a [[model-routing]] cost-ladder wants: send cheap/mechanical/bulk work (refactors, codegen, test scaffolding — the delegate skill's domain) to a DeepSeek-V4-Flash / Qwen-3.6 / Kimi-class tier at ~1/20th the cost, reserve Opus 4.8 / GPT-5.5 for architecture, personality, and final review. This mirrors the lab's existing Opus-orchestrator-+-Sonnet/Codex-worker pattern, now extendable down to open weights. The unified lab/ranker/inspector cost-ladder rig is the natural place to wire and measure this.
2. Provider diversification against availability risk. Fable 5's export-gating is a live lesson: don't pin a critical path to one proprietary super-model. Keeping an open-weight fallback rung (self-hostable, MIT) is now a resilience decision, not just a cost one.
3. Measure, don't trust the press release. Given the benchmark soup, route decisions should ride the lab's own re-runs (bench.blal.pro), not vendor headlines — exactly the discipline NAO already enforces ("raw gpt-5.5 ≠ Alice harness").

Open architectural questions the lab can now interrogate cheaply, because the weights are MIT: [[deepseek-architecture]] (V4's hybrid CSA attention), [[mixture-of-experts]] (1.6T/49B-active sparsity economics), [[scaling-laws]] (does the open wave confirm or bend the curve?), and [[fable-5]] (the gated super-tier's place — if any — in Alice's routing).

See also

• [[deepseek-architecture]] — DeepSeek V4's MoE + hybrid-attention internals
• [[fable-5]] — the gated Mythos-class super-tier in depth
• [[model-routing]] — how OpenAlice picks provider/model per task
• [[mixture-of-experts]] — the sparsity that makes 1.6T models affordable
• [[scaling-laws]] — does the 2026 wave confirm the curve?
• [[llm-evaluation]] · [[benchmark-contamination]] — why the numbers above need caveats
• [[agentic-rl-long-horizon-coding]] — Kimi K2.6's 13-hour-session stability
• [[small-language-models]] — Qwen-3.6-35B-A3B / Gemma 4 self-host tier
• [[ai-safety-and-jailbreaks]] — Fable/Mythos safety classifiers + export gating