Anthropic shipped Fable 5 and Mythos 5 to everyone at $10/$50. OpenAI previewed GPT-5.6 Sol, Terra, and Luna at half the price — but only ~20 government-approved partners can touch it. The 17-day gap between the two launches is the whole story of the 2026 AI market.
Share This Report
Copy the link, post it, or save a PDF version.
Two frontier launches arrived seventeen days apart, and the more capable family is the one you can't buy.
On June 9, 2026, Anthropic shipped Claude Fable 5 and Claude Mythos 5 at $10 input / $50 output per million tokens — generally available in the API on day one. On June 26, OpenAI previewed GPT-5.6 Sol, Terra, and Luna at $5/$30, $2.50/$15, and $1/$6 — undercutting Anthropic's flagship by half — but locked the whole family behind a limited preview that only about 20 government-vetted partners can access. The cheaper, better-packaged release is the one most teams cannot put into production.
That inversion is the story. This is not a "which model is smarter" post. It is a read on what each launch tells you about where the market is going, and what a buyer should actually do about it.
TL;DR
- Claude Mythos 5 — Anthropic's flagship, #1 on BenchLM's blended score (99). $10 / $50, 1M+ context, fully available. The capability-and-trust play.
- Claude Fable 5 — Same $10 / $50 price, blended score 95, positioned as the high-throughput sibling. Anthropic flattened its lineup into two names at one price.
- GPT-5.6 Sol — OpenAI's flagship at $5 / $30 (half of Mythos). New "max" reasoning and "ultra" sub-agent modes. Limited preview only.
- GPT-5.6 Terra / Luna — $2.50 / $15 and $1 / $6. A price for every workload tier.
- The catch — All three GPT-5.6 models cleared OpenAI's "High" cyber threshold; the rollout is gated under the June 2, 2026 federal frontier-model framework. Anthropic gated Mythos-class cyber capability the same way a quarter earlier.
Jump to the summary table for the numbers at a glance.
Strip away the model cards and two opposite strategies sit underneath these launches.
Anthropic bet on one capability tier, one price, available to everyone. There is no mini, no nano, no price ladder. Fable 5 and Mythos 5 carry the same $10 / $50 sticker; the only choice you make is capability versus throughput. Full availability is itself the product feature — Anthropic is selling certainty.
OpenAI bet on a laddered family, priced aggressively, gated on access. Sol for the hardest problems, Terra for high-volume business work, Luna for cheap everyday tasks — a SKU for every budget, each undercutting Anthropic. The cost of being first past a new regulatory tripwire is that you ship into a controlled preview instead of a launch.
The thing that makes them rhyme: both labs gate their most capable models on cyber capability. That is the single most important market signal of 2026. The frontier is now fast enough that shipping is a governance decision, not just a product one. Capability has stopped being the scarce thing. Trust, access, and cost governance are the new frontier.
Anthropic retired the Opus/Sonnet naming for its "5" generation and shipped two models: Mythos 5, the flagship, and Fable 5, its sibling. Both are reasoning models, both carry a 1M+ context window, both are priced at $10 / $50 per million tokens, and both went generally available in the API on launch day.
That GA matters because of what preceded it. In April, Anthropic announced "Mythos Preview" — a frontier model it then declined to ship, citing safeguards that did not yet exist, and routed through Project Glasswing, a controlled-access program with twelve launch partners and forty more organizations doing defensive security work. June's Mythos 5 is the productized, safeguarded descendant of that arc. Anthropic spent a quarter building the deployment scaffolding, then shipped to everyone.
The strategic read: Anthropic collapsed choice. One premium price, two capability points, no access friction. Simplicity as a trust signal.
OpenAI previewed three models at once:
Alongside them came new machinery: a max reasoning mode for deeper inference, an ultra mode that coordinates sub-agents on complex tasks, reworked prompt caching (cache writes billed at 1.25x the input rate, a 30-minute minimum cache life, explicit cache breakpoints), and a planned Cerebras deployment running up to roughly 750 tokens per second in July.
The catch is availability. GPT-5.6 launched as a limited preview to about 20 organizations, shared with the U.S. government before wider release, with general availability promised "in the coming weeks." OpenAI's own framing is that these restrictions "shouldn't be the norm" — a lab visibly uncomfortable with the gate it is complying with.
The strategic read: OpenAI is fighting on price and packaging, and absorbing access friction as the cost of being first through a regulatory tripwire.
| Claude Mythos 5 | Claude Fable 5 | GPT-5.6 Sol | GPT-5.6 Terra | GPT-5.6 Luna | |
|---|---|---|---|---|---|
| Role | Flagship | High-throughput sibling | Flagship | Balanced mid-tier | Fast / low-cost |
| BenchLM blended score | 99 | 95 | preview — not yet ranked | preview — not yet ranked | preview — not yet ranked |
| Price (input / output $/M) | $10 / $50 | $10 / $50 | $5 / $30 | $2.50 / $15 | $1 / $6 |
| Context window | 1M+ | 1M+ | 1M | 1M | 1M |
| Reasoning | Yes | Yes | Yes (max / ultra) | Yes | Yes |
| Availability | GA | GA | Limited preview | Limited preview | Limited preview |
| Released | Jun 9, 2026 | Jun 9, 2026 | Jun 26, 2026 | Jun 26, 2026 | Jun 26, 2026 |
The GPT-5.6 rows show "preview — not yet ranked" for a reason worth stating plainly: OpenAI shipped preview evaluation figures, not a cross-comparable benchmark suite, so BenchLM tracks the models but does not yet rank them. We are not going to invent scores to fill a table. The rest of this post is explicit about which numbers are independently anchored and which are vendor preview claims.
Anthropic published a comparable suite. Mythos 5 sits at the top of BenchLM's blended leaderboard at 99, with Fable 5 at 95, and Mythos leads the verified agentic and reasoning categories. Those numbers are still partly vendor-conditioned, but they are mapped against the same benchmarks every other ranked model uses.
OpenAI shipped preview evaluations only, and they are narrow by design. The figures it chose to lead with:
Treat every one of those as a vendor upper bound until independent evaluations land. But notice which numbers OpenAI published — agentic and cyber — and which it withheld. That selection is the real signal.
Two things are happening at once. Classic benchmarks are saturating: when every frontier model scores in the high 90s on GPQA or MMLU, the benchmark stops discriminating, and labs stop leading with it. At the same time, cyber-capability scores have become the regulated, headline metric — the number governments now use to decide whether a model is a "covered frontier model." So launches increasingly read like security disclosures: a few agentic and cyber figures up front, the rest deferred to general availability. This is why GPT-5.6 is tracked but not ranked on BenchLM, and why the most honest comparison today is about strategy and price, not a leaderboard row.
Mythos 5 and Fable 5 both cost $10 / $50. GPT-5.6 Sol is $5 / $30 — about 2x cheaper on input and 1.7x cheaper on output than Anthropic's flagship. Terra is $2.50 / $15. Luna is $1 / $6 — a 10x input spread from Anthropic's flagship to OpenAI's cheapest tier.
Anthropic is selling certainty: one price, full access. OpenAI is selling a dial: pick your point on the cost/quality curve.
Sticker price is the wrong unit for agents. Consider a single coding-agent task that fans out across 15 tool calls, each averaging 50K input and 10K output tokens — 750K input and 150K output in total. Here is what that one task costs:
| Per agentic task (750K in / 150K out) | Cost |
|---|---|
| Claude Mythos 5 / Fable 5 ($10 / $50) | $15.00 |
| GPT-5.6 Sol ($5 / $30) | $8.25 |
| GPT-5.6 Terra ($2.50 / $15) | $4.13 |
| GPT-5.6 Luna ($1 / $6) | $1.65 |
One task on Mythos costs nearly what nine tasks cost on Luna. Multiply by the thousands of agent runs a production system fires per day and the tiering stops being cosmetic.
This is where the market data reframes the whole conversation. Token prices have fallen roughly 280x over two years — and yet total enterprise AI spend rose about 320% over the same period. The reason is agents: an agentic workflow reasons iteratively, calls tools, verifies, and self-corrects, triggering 10 to 20 LLM calls per user task and consuming 5–30x the tokens of a single chatbot turn (Gartner). Cheaper tokens, far more of them.
The implication for these five models: the per-token sticker matters less than (a) how many tokens a task burns and (b) cache economics. OpenAI's reworked caching — 1.25x cache writes, a 30-minute minimum cache life, explicit breakpoints — is arguably a bigger cost lever than the headline rate for any agent that re-reads a large system prompt or codebase every turn. Anthropic's counter is prompt caching that claws back up to 90% on repeated input. For a long-running agent, the caching design can move the bill more than the sticker does.
Current frontier pricing is widely understood to be a subsidized floor. Analysts expect price normalization within 12–24 months, and OpenAI is reported to lose money on inference at current rates. A team locking a multi-year agent architecture to today's prices is building on a moving foundation.
The practical takeaway is not "pick the cheapest." It is design for model portability and token efficiency — own your eval harness, abstract the model behind a router, and treat any single flagship as swappable — so the next price move or access gate is a config change, not a rewrite.
For workloads that are steadier than the per-task example, here is cost = input × in_price + output × out_price at three monthly volumes:
| Monthly volume (input / output) | Mythos 5 / Fable 5 | GPT-5.6 Sol | GPT-5.6 Terra | GPT-5.6 Luna |
|---|---|---|---|---|
| 1M / 200K | $20.00 | $11.00 | $5.50 | $2.20 |
| 10M / 2M | $200.00 | $110.00 | $55.00 | $22.00 |
| 100M / 20M | $2,000.00 | $1,100.00 | $550.00 | $220.00 |
Cache hits move every row, and unevenly — Anthropic's up-to-90% input discount and OpenAI's new breakpoint caching both apply on repeated context, so a cache-heavy agent narrows the gap the table shows. Use the cost calculator with your real input:output mix before committing. At hobby scale the dollar gaps are a rounding error; at 100M monthly input the gap between Mythos and Luna is most of an engineer's salary.
Here is the thread that ties the two launches into one market.
On June 2, 2026, the White House issued an executive order on advanced AI innovation and security. The framework is deliberately voluntary — it explicitly does not create a licensing, preclearance, or permitting regime. But it asks developers to give the federal government up to 30 days of early access to "covered frontier models" before wider release, and it directs agencies to build a classified benchmarking process, with the threshold for what counts as "covered" set by the Director of the NSA on the basis of cyber capability.
GPT-5.6 is the first marquee model through that tripwire. All three models cleared OpenAI's "High" internal cyber threshold; the ~20-partner, government-shared preview is the framework playing out in public. OpenAI complied and said out loud that it thinks the restriction shouldn't be the norm.
Anthropic got there first, by a different route. Project Glasswing in April was a lab-initiated controlled-access program for cyber-capable models — give defenders coordinated early access before the same capability proliferates. By June, Anthropic had productized the safeguarded version as Mythos 5 and shipped it to everyone. Two paths, one destination: capability now outpaces the safety and governance scaffolding, so access becomes staged.
The market consequence is concrete. "Can I actually deploy this?" has moved from footnote to frontline procurement question. A model you can't get is worth zero to a roadmap, no matter its benchmark. In that light Anthropic's full-availability posture is a competitive product feature, not just a philosophy — and OpenAI's price advantage is partly theoretical until GA opens the gate. Expect more launches to bifurcate into "partner preview" and "GA" phases, and expect cyber-eval scores to become standard launch collateral.
Five directional reads, each with a data anchor.
Spend is migrating to Anthropic. By recent enterprise-spend reporting, Anthropic now captures around 40% of enterprise LLM spend, up from roughly 12% two years ago, while OpenAI has slipped from about half the market to roughly a quarter. Read the two launches as moves in that share war: Fable and Mythos 5 defend a trust-and-availability lead; GPT-5.6's price ladder is built to win spend back on cost.
Agents are the demand driver, not chat. Analysts expect roughly 40% of enterprise applications to embed task-specific agents by the end of 2026, up from under 5% in 2025; Gartner reports 80% of apps shipped or updated in Q1 2026 already embed at least one. Executives are funding it — 88% plan to raise AI budgets for agentic initiatives — and forecasts put enterprise agent spend around $1.4 trillion by 2027. Both launches are explicitly agent-shaped: ultra sub-agents, agentic coding, long-horizon tool use.
Procurement is shifting from model to platform. Buyers increasingly evaluate fine-tuning infrastructure, data-governance controls, and inference economics over raw base-model capability. "Which model is smartest" is being replaced by "whose platform de-risks my deployment."
Reliability beats raw intelligence. Reliability is the most-cited barrier to agent adoption, and only about one in five organizations has a mature governance model for autonomous agents. Trust-scoring frameworks — accuracy, security, explainability, and more — are emerging as procurement gates. A model that is two points smarter but flakier loses the deal.
Specialization presses up from below. Fine-tuned, domain-specific models often beat general frontier models on narrow tasks — cheaper, and runnable where data can't leave the building. So the frontier APIs are squeezed from the bottom by specialization even as cyber-gating squeezes them from the top.
The expectations have inverted since 2024. Capability was the headline then; deliverability is the headline now. A buyer's checklist, with how each family answers it:
A reminder on the numbers: every GPT-5.6 figure here is a vendor preview claim — an upper bound until general availability and independent evaluations land. That is exactly why BenchLM tracks the models but does not yet rank them. Check the live leaderboard and the model pages for current verified standings.
The gap between these two launches is the market in miniature. Anthropic bets that trust plus availability plus capability wins the enterprise. OpenAI bets that price laddering plus ecosystem reach wins it back — once the governance gate opens.
The deeper signal sits underneath both: in 2026, capability stopped being the scarce thing. Deliverability — can you get it, trust it, afford it at agent scale, and govern it — is the new frontier. The lab that productizes deliverability fastest wins the next eighteen months, not the one with the highest benchmark.
So the open question is the one to watch through the back half of the year: when GPT-5.6 reaches general availability and independent evals arrive, does the price ladder pull spend back — or has Anthropic's availability-as-a-feature posture already redefined what "frontier" means to a buyer?
These rankings update with every new model. We send one email a week with what moved and why.
Share This Report
Copy the link, post it, or save a PDF version.
On this page
Which models moved up, what’s new, and what it costs. One email a week, 3-min read.
Free. One email per week.
Four frontier LLMs now advertise 1M+ tokens. DeepSeek V4 Pro's 384K output changes generation workflows. Gemini leads effective-context evals. Here's the real comparison.
Three frontier flagships launched in eight days. DeepSeek V4 Pro undercuts GPT-5.5 by ~9x on output price under MIT license. Here's how they compare on benchmarks, cost, and real use.
Current Anthropic Claude API pricing from official model pages and the Claude Opus 4.7 launch announcement, including prompt caching, batch discounts, and current long-context notes.