Here’s what this call rests on. A company talking about its own model is marked “Official.”
Artificial Analysis / Independent test / Sep 22
Opus 5.5 leads the Intelligence Index and professional-work evaluations; terminal coding is level with Astra. These are model-and-harness tests, not a test of every consumer plan.
Artificial Analysis / Independent test / Sep 21
Independent xhigh testing finds strong professional-work and coding gains, mixed long-context and automation results. Native Grok Build tests are reported separately from the standardized model benchmark.
Anthropic / Official / Sep 22
The latest Opus release adds work and communication improvements. Vendor scores and early-access testimonials are treated as launch evidence, not independent replication.
SpaceXAI / Official / Sep 21
The release verifies Grok 4.7 in Grok Build and developer tools. It does not establish identical access or reasoning limits in every Grok chat subscription.
Google / Official / Sep 2
Google confirms 3.8 Flash for AI Pro and Ultra. The free and AI Plus plans are not assumed to include the same flagship access.
OpenAI / Official / Sep 23
Go is the first paid tier; Pro includes Astra reasoning. Free has limited research and desktop Work access. Live rollout notes take precedence over older model rows.
Anthropic / Official / Sep 23
Pro includes Opus, Research and Claude Code. Max 20x increases capacity. Free includes Sonnet and Haiku, but not Opus or the dedicated Research feature.
Google / Official / Sep 23
Free lists 3.6 Flash, limited 3.1 Pro and Deep Research. AI Plus expands usage; Ultra’s largest option adds more capacity and access to advanced features.
Artificial Analysis / X / Sep 22
The indexed post announces the new Intelligence Index leader. Full post access was unavailable; the linked evaluation is the substantive evidence, not a second independent vote.
Matthew Groff / Article / Sep 6
A first-hand account finds both useful and highlights workflow and usage-limit tradeoffs. It predates Opus 5.5, so it provides context rather than a current winner.
Google / Official / Sep 23
The September 23 rollout adds productivity and creative connections, including Airtable, Linear, Adobe and Webflow. The announcement does not specify identical access for every plan or independently demonstrate task reliability.
SpaceXAI / Official / Sep 23
The individual comparison lists Lite below SuperGrok, then Plus and Heavy. The retrieved table omits some entitlement cells. The separate all-plan Build announcement and 4.7 release support Build access; they do not specify identical allowances.
SpaceXAI / Official / Aug 19
Web/mobile Build expanded to every plan on August 19. Together with the 4.7 release, this supports a public Build path for Lite and Heavy; the free entry point is explicitly a trial.
Andon Labs / Independent test / Sep 24
In six runs per model of a simulated year managing a vending machine, Astra averaged $15,515, Sol $14,428, Grok 4.7 $10,537 and Opus 5.5 $9,235. This narrow API agent simulation favors OpenAI on its profit objective; it does not measure general task completion or consumer-plan access.
Vals AI / Independent test / Sep 22
Opus 5.5 leads the Vals Index and several coding and professional benchmarks. Most runs use max effort; some refusals route to Opus 5 or 4.8, and counting fallback-assisted tasks as failures lowers Terminal-Bench 4.0 from 61.62% to 53.54%. These are model tests, not Max-plan app tests.
Anthropic / Official / Sep 27
Sonnet 5.5 launched September 28. Anthropic reports large coding and professional-work gains over Sonnet 5, faster output, and clearer writing. Its headline Terminal-Bench 4.0 result is a vendor run; higher effort and API harness results are not consumer-plan tests. Independent post-launch replication is not yet published.
OpenAI / Official / Sep 28
OpenAI announced Pro 500, GPT-6.1 Sol in Work/Codex, Codex Cloud and Dots. Sol performance and agent benefits are vendor claims; rollout and limits apply.
OpenAI / Official / Sep 28
GPT-6.1 Sol rolls out to eligible paid Work/Codex plans, not regular Chat. Astra Ultrafast is available on Pro 500, using included allowance before any optional credits; it is a speed tier, not a stronger model.
OpenAI / Official / Sep 28
Dots gradually roll out to US Pro subscribers. They are Astra-powered persistent agents; no independent reliability evaluation or stable post-preview usage terms are published.
Artificial Analysis / Independent test / Sep 27
At max effort, Sonnet 5.5 scores 56 on the Intelligence Index and 64% on Terminal-Bench 4.0, near Opus 5.5 on work tasks. It used unusually many tokens; lower efforts lagged, default fallback was enabled, and some prerelease structured-output runs had a bug.
Vals AI / Independent test / Sep 27
Sonnet 5.5 scores 69.22% on the Vals Index and leads Code Migration and Vibe Code. Max effort and server-side Sonnet 5 fallback affect interpretation; counting fallback-assisted tasks as failures lowers its Vals Index to 68.89%.
Artificial Analysis / Independent test / Sep 28
At max effort Sol 6.1 scores one point below Astra on AA Intelligence Index; coding-agent gains vary by effort, with xhigh above Astra on that index. This is a model/harness comparison, not a Pro 500 subscription test. Cost-per-task is API cost, not included-plan value.
Vals AI / Independent test / Sep 28
At max effort Sol 6.1 records 61.15% on the Vals Index (7/30), 55.05% on Terminal-Bench 4.0 (5/31), and 65.12% on Code Migration (4/70). This differs from AA’s near-Astra aggregate; harnesses and settings are not interchangeable.
Google / Official / Sep 29
Argon is initially limited to trusted cyber defenders in the Fairwind Program. Google says wider developer, enterprise and consumer access comes later, starting with paid API and Google AI Ultra. Its coding, work and creative-writing benchmarks are vendor claims until independently checked; no current US individual plan is confirmed to have Argon.
Vals AI / Independent test / Sep 29
Vals ranks Argon first on Vals Index 2.1 at 68.90% ±0.97, with high reasoning effort; it is second on Code Migration and Vibe Code, fifth on Terminal-Bench 4.0. This tests a model through an evaluation harness, not access or outcomes in a Google AI Ultra subscription.