Could Mark Zuckerberg’s Open-Weight Coding Gambit Slash OpenAI and Anthropic’s Valuations?

Published and source-checked August 10, 2026.

Key takeaways

• Meta’s move is real but incomplete. Muse Glimmer is available now; the open-weight Muse Spark 1.2 release is still a promise.
• Coding is the best place to attack closed-model economics because demand is large, token use is heavy and results can often be checked with tests, builds and code review.
• Meta can tolerate thin or zero direct model margins because AI also improves its advertising business, products, devices and developer ecosystem.
• OpenAI and Anthropic still own serious moats: better frontier performance in important workloads, trusted products, enterprise controls, distribution and fast-growing customer bases.
• A valuation shock becomes plausible if open models become close enough in quality that developers route most work away from premium APIs. It does not require Meta to build the single best model.

What Meta actually released

On August 10, Meta released Muse Glimmer, a dense model with roughly 29.6 billion parameters and a context window of at least 131,072 tokens. Meta describes it as an agent model for local tool use, coding, multimodal work and failure recovery. The weights are on Hugging Face under Apache 2.0.

The local-deployment details matter more than the round number. Meta says a four-bit version fits in less than 20GB, within the reach of common 24GB or 32GB systems. Its tests reached 233.4 tokens per second on an Nvidia RTX 5090 and 37.8 to 50.2 on high-end Apple silicon. Those are vendor measurements, not independent guarantees, but they describe a model that can run without sending every file to a remote API.

Glimmer also looks capable for its size. Meta reports 51.2% on SWE-Bench Pro, 76% on SWE-Bench Verified and 51.7% on Terminal-Bench 2.1. It beats Meta’s listed Gemma 4 31B result on all three and roughly matches or trails Qwen 3.6 27B depending on the test. The honest reading is “strong small model,” not “Claude or Codex killer.” Meta’s own table shows weaknesses, and the company recommends system-level guardrails for deployment.

The larger bet is Muse Spark 1.2. Meta launched Spark 1.2 and Muse Code on August 5. The coding-focused model was trained with the Muse Code terminal agent for long repository tasks and sustained tool use. It is available through Meta’s API today. Meta says weights for a version will arrive “in coming weeks,” according to Axios.

That wording deserves restraint. We do not yet know the open version’s final license, size, hardware requirements, benchmark parity or release date. Until the files arrive and independent testers reproduce the results, Spark 1.2 is an announced catalyst rather than an established market fact.

There is a terminology wrinkle. Meta calls Glimmer open source, and Apache 2.0 grants broad rights over the weights. The Open Source Initiative’s AI definition, however, also expects information and code needed to study how the parameters were produced. Meta lists training-data categories but not a reproducible dataset. “Open-weight” is more precise—and commercially consequential.

Why the starting valuations leave little room for an ordinary outcome

As Kingy.ai’s comparison of their economics makes clear, OpenAI and Anthropic are not valued like normal software companies.

Sources: OpenAI’s March 31 funding announcement and Anthropic’s May 28 funding announcement.

These are post-money private funding marks, not continuously traded prices. Annualized revenue is a snapshot, not audited full-year revenue. Even so, the ratios reveal what investors paid for: rapid growth, durable pricing power, improving margins, frontier leadership and a large market.

Open weights attack the pricing-power part of that package. Suppose, only as an illustration, that investors cut each company’s revenue multiple in half while its current run rate stayed unchanged. OpenAI’s mark would fall from $852 billion to $426 billion. Anthropic’s would fall from $965 billion to $482.5 billion. No bankruptcy, customer exodus or technological collapse is required. A lower belief about future margins can do the damage.

That is an illustration, not a forecast. Revenue could grow fast enough to offset a lower multiple. It simply shows how “massive” compression is possible even if both companies remain important.

Coding is the pressure point

Coding has become the commercial proving ground for agents. It consumes huge amounts of context and output tokens but provides unusually clear feedback. Code compiles or it does not. Tests pass or fail. A patch can be reviewed, benchmarked and reverted.

That makes buyers more willing to spend, and more willing to switch.

Anthropic said in February that Claude Code had exceeded $2.5 billion in run-rate revenue, more than double its level at the start of 2026. OpenAI said in June that Codex had surpassed five million weekly active users. Kingy.ai’s hands-on framework for Claude Code versus Codex reaches the right practical conclusion: these are full workflows with terminals, IDEs, cloud workers, review systems and permissions, not interchangeable chat boxes.

Coding agents are expensive to operate. A long task may read a large repository, call tools hundreds of times, reject several approaches and retain context for hours. Small token-price differences multiply. The Associated Press recently documented U.S. developers and companies adopting cheaper open Chinese models, with agent workloads magnifying the difference.

This is Meta’s opening. Its model need not beat Claude Opus or GPT-5.6 on every hard problem. It can take the broad middle: code search, test generation, routine migrations, documentation, triage, small fixes and background subagents. A router can reserve a premium model for the hardest planning or recovery steps.

If an open model handles 70% of a coding agent’s tokens and a premium model handles the remaining 30%, the proprietary lab may still provide indispensable intelligence while collecting far less revenue from the task. That mix is dangerous to valuations built on both high usage and high unit economics.

Meta does not need to make money the same way

OpenAI and Anthropic primarily sell intelligence through subscriptions, APIs and enterprise products. Meta’s base is advertising. In the second quarter of 2026, it reported $60.8 billion in revenue, including $59.36 billion from ads, and 3.60 billion daily users across its apps.

That gives Zuckerberg two weapons.

First, Meta can fund a price war from a profitable business that benefits when AI improves recommendations, ads, messaging and creator tools. It expects $130 billion to $145 billion of 2026 capital expenditure, an amount few standalone labs can match without repeated fundraising.

Second, Meta can treat models as a complement to products it owns. Cheap intelligence can improve WhatsApp, Instagram, Facebook, glasses and advertising tools. A large ecosystem around Meta’s formats and runtimes may be worth more to it than premium margins on every call.

This is not charity. Zuckerberg’s “Future is for Everyone” essay presents open models as a check on centralized power, but the policy fits Meta’s position. Open weights weaken rivals built around closed APIs while strengthening a company that monetizes billions of users elsewhere. Principle and self-interest can point in the same direction.

How the valuation compression could happen

The bear case for OpenAI and Anthropic works through five connected effects.

1. The premium-price umbrella breaks

When a downloadable model is good enough, cloud hosts compete to serve the same weights. Hardware vendors optimize them. Quantization improves. Startups offer lower-cost endpoints. That competition pushes inference pricing toward compute cost, leaving less room for model-owner margins.

2. Enterprises gain a credible self-hosted option

Companies with sensitive source code care about data location, retention, latency and vendor dependency. A strong open model can run inside their own cloud account or network, be fine-tuned on internal conventions and remain available even if a provider changes pricing or policy. Self-hosting is not free, but control can justify the operational burden for large users.

3. The agent system becomes more valuable than the model

Muse Code, Codex, Claude Code, OpenCode and other agents combine models with permissions, memory, tools, sandboxes and review loops. If customers can swap the model without rebuilding the workflow, it becomes a replaceable component. Value migrates toward the interface, enterprise control plane, cloud platform and developer community.

OpenAI and Anthropic know this; both are building broad agent products. Meta’s attack would not remove the application layer. It would force closed labs to prove that their integrated product is worth more than an open model inside a capable third-party agent.

4. Community work compounds Meta’s release

Open weights invite people to produce quantizations, fine-tunes, deployment recipes, evaluations and hardware-specific kernels. Meta pays for the base training run; the ecosystem performs much of the adaptation. Each improvement opens another niche or device.

5. “Good enough” expands faster than the frontier

Frontier leadership still matters for the hardest tasks. Most paid work is easier. As smaller models improve, yesterday’s premium capability moves into a cheaper tier. Closed labs must spend to create a fresh gap while the open ecosystem absorbs the previous one. They can grow impressively and still produce worse economics than today’s valuations imply.

Why Anthropic may be more exposed—and better defended

Anthropic’s coding success makes it the clearest target. Claude Code created a powerful developer brand and an entry into enterprise workflows. A cheaper open stack challenges that wedge directly.

Yet Anthropic’s $47 billion run rate is nearly twice OpenAI’s disclosed $24 billion annualized pace, while its valuation is only about 13% higher. On that snapshot, Anthropic’s multiple is less stretched. It also offers Claude across major clouds and has deep enterprise relationships. Customers buy reliability, security, support and product experience alongside tokens.

OpenAI is more diversified. By March, ChatGPT had more than 900 million weekly users and 50 million subscribers, while enterprise revenue exceeded 40% of the total. Codex is moving beyond software development into knowledge work. Those businesses reduce its dependence on coding-model rent.

OpenAI’s disclosed multiple is higher, however, and its infrastructure commitments require enormous future cash generation. If model prices fall faster than usage rises, the market could question how much value belongs to OpenAI’s proprietary research rather than ChatGPT distribution and the surrounding product suite.

The threat differs: Anthropic has more coding concentration but stronger recent revenue relative to its mark. OpenAI has broader distribution but a richer valuation and a larger story to defend.

The case against the Meta thesis

There are strong reasons this gambit may produce pressure without producing a valuation collapse.

Glimmer is not a frontier model. Its footprint is impressive, but Meta’s scores do not put it above the best Claude and GPT systems. One subtle architectural error can erase every token saving.

Spark 1.2’s open release may disappoint. Meta could delay it, release an altered checkpoint, use a restrictive license or require impractical hardware. The market should judge the artifact, not the promise.

Running weights is not running a product. Enterprises need identity controls, audit logs, observability, support, security testing and uptime. Engineers must manage GPUs, updates and capacity. Closed services can win on total cost even when their tokens cost more.

Products create switching costs. Claude Code and Codex sit inside established workflows. Their vendors can bundle models, agents, storage, search and enterprise contracts. An open model may push prices down while the product keeps the customer relationship.

OpenAI and Anthropic can respond. They can cut prices, distill smaller models, improve caching and route tasks across tiers. OpenAI already distributes Apache-licensed open weights. Meta may accelerate the whole market rather than secure a permanent advantage.

Trust cuts both ways. Some companies will prefer local weights. Others may hesitate to build a critical stack around Meta’s roadmap, data practices or shifting priorities. Anthropic’s safety positioning and both rivals’ enterprise commitments have value.

Three scenarios to watch

The signals matter more than launch-day applause. Watch independent cost per completed task, sustained developer traffic, enterprise self-hosting and long-horizon coding results under comparable agent systems. Most of all, watch whether OpenAI and Anthropic can keep revenue growth high while per-unit prices fall.

The verdict

Zuckerberg has found a credible way to attack closed-model economics without defeating every closed model: release capable coding weights, let hosts compete away serving margin, and use Meta’s advertising cash to sustain the fight.

That strategy could cut both valuations sharply if coding becomes a low-margin commodity and investors stop assigning frontier-lab multiples to portable revenue. The risk matters because agents consume heavily, businesses can measure output and both companies have built major growth stories around coding.

But the evidence supports a warning, not a coronation. Glimmer is a strong local model, not a demonstrated frontier winner. Spark’s weights are not public. Claude Code and Codex have scale, trust and workflows that weights alone do not replace.

Meta does not need to kill either company to change what each is worth. It only needs to make proprietary coding intelligence less scarce than investors assumed.

FAQ

Is Muse Glimmer truly open source?

Meta released the artifacts under Apache 2.0, while its repository also carries a separate prohibited-use policy. “Open-weight” is the safest term because Meta has not released everything needed to reproduce the training process under the Open Source Initiative’s definition.

Has Meta released Muse Spark 1.2’s weights?

No. Muse Spark 1.2 is available through Muse Code and Meta’s API. Meta says weights for a version will arrive in the coming weeks, but the final artifact, license and hardware requirements are not yet public.

Would open models make OpenAI or Anthropic worthless?

No. Both companies have products, distribution, enterprise contracts, research teams and infrastructure beyond any one model. The credible risk is lower margins and valuation multiples, not automatic business failure.

Why focus on coding?

Coding agents consume many tokens, produce measurable outputs and sit close to enterprise spending. They are also easier to route across models than a tightly integrated consumer assistant, which makes price and deployment flexibility unusually important.

What would prove the thesis wrong?

The thesis weakens if Spark’s open weights do not arrive, perform poorly outside Meta’s own agent, remain costly to run, or fail to win sustained enterprise and developer adoption. Continued rapid growth at OpenAI and Anthropic despite lower market prices would also show that expanding demand matters more than commoditization.

Methodology and disclosure

This is scenario analysis, not investment advice. Kingy.ai did not run a new hands-on test for this article; model claims and benchmark figures come from the linked vendor materials and are labeled accordingly. Private valuations and run-rate revenue are point-in-time company disclosures, not audited public-market measures. The article was source-checked on August 10, 2026.

Selected sources

• Meta AI Research: Introducing Muse Glimmer
• Muse Glimmer model card and weights
• Meta AI Research: Introducing Muse Code and Muse Spark 1.2
• Mark Zuckerberg: The Future is for Everyone
• Axios: Zuckerberg’s AI manifesto and Meta’s open-weight plans
• OpenAI funding and operating metrics, March 31, 2026
• Anthropic Series H funding and revenue metrics, May 28, 2026
• Meta second-quarter 2026 results
• Open Source AI Definition 1.0