Anthropic, Google, and Meta Shipped New Models in the Same Week

OpenAI spent the week owning the narrative with GPT-6 Astra. We already covered that rollout, brakes and all. The rest of the industry did not wait. CNET’s roundup is the right frame: Anthropic, Google, and Meta put new weights on the table in the same few days.

36Kr, translating a CSDN recap dated September 3, puts Anthropic first on the calendar: Fable 5.1 and Mythos 5.1 on September 1. Google followed with Gemini 3.8 Flash and Gemini 3.8 Flash Cyber. Meta shipped Muse Spark 1.3. If you only read the Astra thread, you missed the part that changes what you actually subscribe to.

Three labs, one calendar week

Version numbers in 2026 land in clusters because nobody wants to go a week without a press cycle. This cluster is denser than usual. Anthropic’s own post pitches Fable 5.1 and Mythos 5.1 as the most advanced models it has for coding and knowledge work, with a research demo attached so the scientific-progress slide has a screenshot.

Google’s drop is a Flash pair, which usually means “cheaper, faster, maybe a cyber SKU.” Meta’s Muse Spark 1.3 is the one trying to sound like a coworker instead of a completion engine.

You do not need a new bake-off spreadsheet on day one. You need to know which of these is generally available, which is a gated twin, and which one will quietly drain tokens if you leave reasoning at max.

The clustering is also a procurement headache. Finance sees three “new SOTA” emails and assumes you must switch. You must not. A week with four version bumps is a week to freeze defaults and read the changelogs. The team that migrates on Wednesday is the team that spends October explaining a broken tool call.

Our Claude vs Gemini product piece was about the shape of the products, not these version strings. Treat this as a news overlay on that review, not a replacement.

Fable vs Mythos is a policy split, not two brains

Anthropic’s sentence is easy to miss and it is the whole release: Fable 5.1 and Mythos 5.1 are the same model with different safeguards. Fable is generally available. Mythos is trusted-access only, with the extra rails aimed at cybersecurity and the life sciences.

That is not “Mythos is smarter.” That is “Mythos is allowed to do work we will not sell in the chat box.” If you are not in the trusted-access program, arguing about Mythos 5.1 on social media is fan fiction.

CNET adds the cost note that matters for people who already pay for Fable: at low or medium effort, Fable 5.1 is supposed to match Fable 5 at a lower bill. If that holds, the upgrade is not a new personality. It is a cheaper way to get the old one, with a higher ceiling if you turn effort up.

36Kr cites Artificial Analysis’s Comprehensive Intelligence Index with Fable 5.1 on top at 66. Indexes are not gospel. They are a way to say the lab did not ship a dud. The same recap warns that at maximum intensity the model eats tokens. That matches the “leave effort on max and then complain about the invoice” failure mode every agent stack hits.

If your job is writing code in a repo, Fable 5.1 is the one you can actually type into. If your job is a cyber range or a wet lab, Mythos is a procurement conversation, not a model picker in the UI.

Same weights, different rails also means your evals will not transfer if you ever do get Mythos. A refusal on Fable is not a bug. A tool call that Mythos will run and Fable will not is the product. Budget time for that split instead of assuming a shadow deployment will “just be smarter.”

The Venus map and what “research” means here

Anthropic’s scientific teaser is specific enough to quote. Fable 5.1 trained a neural network to build a high-resolution elevation map covering a third of Venus, using Magellan radar images more than thirty years old and an existing map of one-fifth of the planet. The new map, Anthropic says, resolves details down to two to three kilometers.

That is a bounded inverse problem with public data. It is not a paper that overturns planetary science. It is also not “the model wrote a poem about Venus.” If you are going to claim research capability, a map with a resolution number is the right kind of claim: checkable, ugly, and tied to old NASA bits.

Do not confuse this with K2 Horizon’s “open frontier” story from last week. K2 Horizon was about weights you can actually download. Fable 5.1 is a hosted model that did a geophysics homework assignment on a press page. Different axes.

The useful takeaway for a working engineer is smaller: these models will keep showing demos that look like science because science has files. Your own “research” workload is closer to that Venus job (messy inputs, a numeric output, a need to not hallucinate units) than to a chat about the meaning of AGI.

Gemini 3.8 Flash is the cheap lane

36Kr’s numbers: Gemini 3.8 Flash scores 59 on the same index at highest reasoning intensity, and you can turn thinking down to cut the cost of a single task. They call it a workhorse. That is the honest product role. You do not buy Flash to win a leaderboard screenshot. You buy it to run the 80 percent of calls that should never have touched Fable.

CNET lists Gemini 3.8 and 3.8 Flash Cyber as Google’s half of the week. The Cyber SKU is the now-standard pattern: a slightly more locked-down twin for customers who have to write “cyber” on a purchase order. It is not automatically Mythos. Anthropic’s trusted-access bar and Google’s Flash Cyber badge are different documents.

If you already standardized on Gemini for volume, 3.8 Flash is the version string to pin and load-test. If you standardized on Claude for hard refactors, do not flip the default because a Flash model gained a point. Run the jobs you actually pay for.

Latency and price still beat index score for classification, extraction, and “summarize this ticket.” Put Flash there. Put Fable 5.1 on the jobs where a wrong API call costs more than the tokens.

Muse Spark 1.3 tries to talk before it acts

CNET’s Meta section is the behavior change, not the benchmark. Muse Spark 1.3 is trained to ask clarifying questions when a prompt is vague, and to confirm actions in agentic workflows. It is available in Muse Code and Meta’s Model API.

36Kr adds the efficiency pitch: fewer tool calls, fewer tokens, better hold on long tasks. That is the opposite of the Fable-at-max-intensity problem. Meta is trying to win the agent loop by doing less.

Asking a clarifying question is polite in a chat window and lethal in an unattended cron. If you wire Spark 1.3 into a headless agent, you need a policy for “the model asked a question and nobody is there.” Either fail closed, or inject a default. Do not assume the confirm-before-act training will no-op when tool_choice is on.

Meta needed a model people mention in the same sentence as Claude and Gemini. Spark 1.3 is that attempt. Availability in Muse Code means it is a product surface, not a research dump. Try it on a long refactor with messy requirements. That is the scenario the clarifying-question training was built for. Do not try it first on a silent batch job.

Watermarks after August 2

Buried in Anthropic’s post is a policy that will outlive the version number. Models released after August 2, 2026, watermark their text. The watermark is a numeric signal you can only read with a detection API. Anthropic says it is invisible otherwise, does not carry user or conversation metadata, and does not dent quality.

If you publish model output as your own without saying so, a detector now has a better chance of catching Claude-after-August-2 than it did last spring. If you are a university or a newsroom that already bought a detector, ask whether it speaks this API.

If you are a company that uses Claude to draft customer email, the watermark is not a privacy leak according to Anthropic. It is a provenance tag. Your legal team still has to decide whether that matters in your jurisdiction. Your engineering team has to decide whether stripping or rewriting output to dodge a watermark is a thing you even want to be in the room for. (You do not.)

Google and Meta’s posts in this week’s secondary coverage did not lead with watermarks. That does not mean their outputs are unmarked. It means Anthropic chose to say it out loud.

How to pick without running another bake-off

Pin versions. Fable 5.1 if you already live in Claude and the low/medium-effort cost claim matters to your bill. Gemini 3.8 Flash if you need a cheaper default and can park Cyber in the security account. Muse Spark 1.3 if you want an agent that asks before it runs rm and you are willing to sit in the loop.

Do not migrate a production tool-calling graph the week three labs ship. Watch the invoices for a few days. Check whether Fable 5.1 at medium effort really matches Fable 5. Check whether Spark 1.3 stalls waiting for answers. Check whether Flash Cyber’s extra rules break your evals.

Astra can remain the loud model. The quiet question is which of these three you will still be calling in November when the blog posts are stale. For most teams that is Flash for volume and one expensive brain for the jobs that fail out loud. This week just renamed both.

One more practical note on the index scores, since 66 and 59 will get screenshotted in Slack. Artificial Analysis is a composite. It is not your latency budget, your SOC-2 questionnaire, or your coding eval. Use it to decide whether a model is in the conversation. Use your traces to decide whether it stays. Fable 5.1 leading the composite and Flash sitting at 59 is consistent with “pay more for the hard jobs.” It is not a reason to delete the cheap lane.