We built OpenDsec: an open-source, self-hosted implementation of the sandbox platform architecture described in DeepSeek’s DSec paper.
It is an independent project—not DeepSeek code or an official release.
A short architecture tour:
Control plane: stateless apiservers handle auth, quotas and placement; Raft-replicated IAM stores identities and project state; watchers aggregate node health; one edge daemon per compute node owns the sandbox lifecycle.
Calls are routed over gRPC/mTLS.
Data plane: each sandbox runs as a Firecracker microVM, container or full QEMU VM. Aether runs inside the sandbox; Chronus maintains shell sessions.
The same runtime API covers exec, files, HTTP, processes and PTYs.
OpenDsec can pause a microVM to a memory snapshot, restore it in ~55 ms, or fork a running VM with memory preserved.
Forks share pages on the same node; cross-node forks pull the snapshot from the source.
Images come from OCI registries, compose without copying layers, and can pack sandbox changes into a new image. Network access is default-deny, with domain/CIDR/port rules.
An E2B-compatible gateway lets existing E2B SDK clients run against it.
On our hardware, a microVM cold-starts in ~0.1 s. One 384-thread node handled 5,120 microVMs in a burst; a two-node cluster handled 8,192 containers with no failures.
These are OpenDsec measurements, not numbers from the DSec paper. #deepseek #sandbox
Anthropic has launched Claude Fable 5.2 and Claude Mythos 5.2, two models built for technical work that is too difficult, ambiguous, or long-running for a typical “daily driver” model.
The simplest way to understand the release is this: Fable and Mythos share the same underlying model, but they are distributed under different access and safety regimes. Fable 5.2 is broadly available and includes strong safeguards around high-risk cybersecurity and research biology. Mythos 5.2 removes some of those restrictions for vetted organizations in Anthropic’s Cyber and Life Sciences Verification Programs.
Anthropic is not positioning either model as the default choice for every prompt. It recommends Opus 5.5 for most coding and knowledge work, with Fable or Mythos reserved for ambitious, open-ended problems. They can also serve as an “advisor” consulted by other models or as an “orchestrator” that delegates simpler work to cheaper systems.
The efficiency story is as important as the benchmark story. Anthropic says the 5.2 models produce output up to 30% faster than their 5.1 predecessors. Standard API pricing is $8 per million input tokens and $40 per million output tokens, down from $10/$50 for Fable 5.1. An optional fast mode can generate tokens up to 2.5× faster at $16/$80 per million tokens.
On Anthropic’s own evaluations, Fable 5.2 scored 49.1% on FrontierCode 2.0 at Xhigh effort, versus 32.4% for Fable 5.1, 45.5% for Opus 5.5, 42.8% for GPT-6 Astra, and 41.1% for Gemini 4 Carbon at Medium effort. On FrontierSWE v2, it reached 69.9%, compared with 56.3% for Fable 5.1 and 62.3% for Opus 5.5. These are vendor-reported results, so they should be treated as directional until independent testing catches up.
The more interesting claim is cost-adjusted performance. At High effort, Fable 5.2 reportedly matched Opus 5.5’s best FrontierCode score for roughly half the cost per task. On Perplexity’s WANDR data-collection benchmark, Fable 5.2 at Medium effort matched Opus 5.5’s top score at about 40% of the cost per task. That suggests the real improvement may be less about a higher ceiling and more about reaching that ceiling with fewer tokens.
Anthropic also highlighted several internal projects. A team of Fable 5.2 agents migrated a critical production service from Python to Rust; the resulting service reportedly became more reliable and started about 26× faster. In another experiment, the model began with an empty folder, coordinated more than 1,000 subagents for eight days, and produced a prototype operating system that boots on a laptop, including storage, networking, windows, login, installation, and built-in Claude functionality.
That operating-system demo matters less as a consumer product than as evidence of orchestration. The hard part of long-running agentic work is not generating one good code snippet. It is preserving intent, dividing work, checking dependencies, testing intermediate results, recovering from mistakes, and keeping hundreds of steps pointed at the same outcome.
Mythos 5.2 is aimed at the dual-use edge. Anthropic calls it its most capable model for cybersecurity engineering, with improvements in discovering vulnerabilities, reproducing them, and producing fixes. Access remains restricted. Fable 5.2 blocks high-risk activities such as exploit generation and many penetration-testing tasks, as well as most research-biology requests. Anthropic says updated classifiers reduced unnecessary blocks by roughly 55% overall versus Fable 5.1—about 60% fewer in cyber and 40% fewer in biology—without weakening the intended safeguards.
The release also shows how model differentiation is changing. The old comparison was mostly “Which model is smartest?” The newer questions are operational: How long can it work without drifting? How well can it coordinate subagents? How much does a correct result cost? What tools can it use? Which capabilities require verified access? And can safeguards become more precise without turning the general model into a refusal machine?
My takeaway: Fable 5.2 and Mythos 5.2 are not really chatbot upgrades. They are infrastructure for projects with long horizons, expensive errors, and a large amount of technical uncertainty. The decisive tests will not be short benchmark prompts. They will be week-long migrations, security investigations, hardware and scientific workflows, and whether the final artifact survives expert review.
Anthropic's new post on agent memory is the clearest architecture writeup they've published. The whole system fits in one sentence: capture in-session, organize offline.
The blueprint is borrowed from the brain. The hippocampus records constantly; consolidation into the neocortex happens offline during sleep. So Claude's memory is split the same way — a fast path that writes notes during the session, and a "dreaming" pass that reorganizes everything between sessions.
Four layers, bottom to top: transcripts (raw event logs, the ground truth), memory notes (salient observations written in-session), representations (wikis of linked pages), and injected context (what gets preloaded into every new session). Dreaming reads the transcripts and works upward through all of it.
The key frame: a memory note is a prediction. Writing down a rule is betting that the pattern will recur. Writing a verified fact is betting a future session can trust it without re-checking. Newer models are dramatically better at this bet.
On CL-Bench, each question runs in a fresh session with a persistent memory directory. Sonnet 4.6 scattered untested guesses across files: 0.330 → 0.364 with memory, barely helps. Opus 5 was the first to close the loop — rules, evidence for the rules, warnings about traps: 0.564 → 0.828.
The Pokémon experiments tested what's worth remembering. Sonnet 3.7 assumed everything might matter: 413 files, 977,930 characters, one badge. By step 10,000 it was spending whole stretches just reorganizing files instead of playing.
Fable 5 wrote 9 files plus a global.md dashboard: current goal, gym plans, its own rules for using memory. Four badges. The test is simple — anything readable off the screen every turn (position, money) is redundant. Only things that vanish unless written down deserve files.
One finding flips common practice: give frontier models the goal, not the method. Telling Opus 5 to "record only failures" cut its memory gain by 0.234. Just saying "a memory folder is available" was enough for it to work out its own strategy. Prescriptive rules weakened performance.
The counterexample: Sonnet 4.6 rarely read its own memory unprompted. Telling it "read memory before answering, then test your theories" nearly doubled its score, 0.364 → 0.641. Weak models need the steering strong models outgrew.
Dreaming corrects memory. In Mt. Moon, Claude wrote down (wrongly) that the PokéCenter was on the east side. It walked out, said "Mt Moon is behind me," fell off a one-way ledge three times, then wrote "I'm going to stop rather than keep looping." Never reached the next city.
One offline pass against the transcript caught the error and rewrote the note. Same run then walked straight back into the cave and reached Cerulean at action 1,935. The transcript is the ground truth the note gets validated against.
For representations, they built wikis: Claude-maintained linked pages over ~350k files / 420M tokens from Stainless, the company they acquired in May. Over 128 sessions, incremental dreaming pushed wiki-only answers from 13% to 48%. Wiki + search: 54% → 61%, vs 53% for search alone, and ~20k tokens saved per question.
Wikis pay off only when questions span multiple documents — 87% vs 73% for search alone on a merger corpus. When answers sit in a single document, text search wins and the wiki adds nothing.
And they hit the Borges failure mode. One wiki grew to 1.77x the original corpus (203KB of wiki for 115KB of docs). A map as large as the territory abstracts nothing. Keeping the injected layer small is half of dreaming's job — Claude Code loads only the first 200 lines / 25KB of MEMORY.md as an index.
The implementation is more prosaic than the theory: memory is just a directory of markdown files with frontmatter and [[slug]] links. No database, no vector store. Dreaming is a batch job — one analyzer agent per transcript, an organizer agent merging duplicates and fixing links, run on cron every night.
Costs are modest: ~$8 and ~10 minutes to dream over 50 long transcripts (~12MB) with Haiku analyzers and a Sonnet lead. Worth noting what they took seriously: transcripts are treated as untrusted input, analyzers get five file tools and no network, secrets are never written to memory, every write is versioned.
Open questions they admit to: when to write memory (in-session capture is immediate but distracts; dreaming is free of attention but not immediately available), how to keep memory from going stale without re-checking everything, and whether transcript search alone should be the foundation.
https://t.co/XWf6n0IV45
Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. https://t.co/uUEN4bNkr9
Your current failures don't matter because success often comes from failures. I think it depends on how you define your failures. You can define failure as an opportunity to do better next time, as long as you keep working on it
Hybrid LLMs promise the best of both worlds: Attention for precise global context, SSMs for cheap recurrent memory. But models like Jamba and Nemotron-H still hard-code one Attention/SSM layout for every request.
Flex asks: why should every prompt take the same path?
The motivation is empirical. Reasoning, QA, and summarization prefer different Attention/SSM mixes—and even requests within the same dataset favor different layouts.
A fixed layout either wastes Attention on easy prompts or removes it where exact context recall matters.
Flex pairs an Attention branch and an SSM branch in every hybrid block. During prefill, a lightweight gate uses the prompt's hidden states to choose one branch per block, creating a request-specific execution path.
Training uses soft routing; inference uses top-1 routing.
The route is frozen after prefill. This is crucial during autoregressive decoding:
• Attention blocks append to their KV cache
• SSM blocks update a bounded recurrent state
Changing branches token by token would invalidate the available history and force recomputation.
Dynamic paths normally break batching, so Flex splits only branch-specific work into compact Attention and SSM sub-batches, runs them separately, then merges them for shared operations such as the FFN.
It also assigns more GPU resources to the heavier branch.
Prefix reuse must match both text and execution path. Flex uses a two-level radix tree: one level for textual prefixes, another for layer-choice histories and caches.
Across 6 datasets, it leads 13/18 quality metrics, cuts TFLOPs/token by up to 67.2% vs Llama, and reaches up to 3.87x vLLM throughput on one H100.
The value of experience is often overvalued.Most breakthrough achievements are made by novices.
Therefore, when hiring, look at the candidate's qualifications and attitude first,and then train them in the skills.
OpenDsec publie une implémentation indépendante inspirée de l’infrastructure de sandbox décrite par DeepSeek : conteneurs, micro-machines virtuelles, quotas, instantanés et contrôle réseau pour faire travailler des agents à grande échelle. Le projet AGPL-3.0 est né cette nuit et n’est pas un produit officiel de DeepSeek.
https://t.co/wmrXLp75sD
We built OpenDsec: an open-source, self-hosted implementation of the sandbox platform architecture described in DeepSeek’s DSec paper.
It is an independent project—not DeepSeek code or an official release.
A short architecture tour:
Control plane: stateless apiservers handle auth, quotas and placement; Raft-replicated IAM stores identities and project state; watchers aggregate node health; one edge daemon per compute node owns the sandbox lifecycle.
Calls are routed over gRPC/mTLS.
Data plane: each sandbox runs as a Firecracker microVM, container or full QEMU VM. Aether runs inside the sandbox; Chronus maintains shell sessions.
The same runtime API covers exec, files, HTTP, processes and PTYs.
OpenDsec can pause a microVM to a memory snapshot, restore it in ~55 ms, or fork a running VM with memory preserved.
Forks share pages on the same node; cross-node forks pull the snapshot from the source.
Images come from OCI registries, compose without copying layers, and can pack sandbox changes into a new image. Network access is default-deny, with domain/CIDR/port rules.
An E2B-compatible gateway lets existing E2B SDK clients run against it.
On our hardware, a microVM cold-starts in ~0.1 s. One 384-thread node handled 5,120 microVMs in a burst; a two-node cluster handled 8,192 containers with no failures.
These are OpenDsec measurements, not numbers from the DSec paper. #deepseek #sandbox
Anthropic just acquired a one-person drug company.
Mithril Medicine: 1 full-time employee, 8 consultants, 11 outsourced labs — coordinated by Claude. Went from idea to positive mouse data for a rare muscular dystrophy in under 18 months.
Anthropic calls it "Pipeline-in-a-Person."
Why buy a one-person company? The deal is small, the logic isn't:
1/ It's a working proof case. A solo founder used Claude to run contracts, logistics, and experiments across 11 labs. Anthropic now owns the story it wants every biotech to believe.
2/ Feedback loops. Anthropic's life sciences head: to build the right models "we need to live it." Own drug programs surface bottlenecks that no API customer report would.
3/ Catch-up. DeepMind's Isomorphic Labs raised $2.1B in May, has ~$3B in pharma deals, and is running its own pipeline. Anthropic built a secret wet lab in the Bay Area (per Reuters), bought Coefficient Bio for ~$400M in April, and put Novartis's CEO on its board.
One caveat: Anthropic says it will stop short of clinical trials to avoid competing with pharma customers. Mithril's "human trials by late 2028" is already the far edge of that line.
@shenshanni@Cara_Mosquito this is the part nobody wants to hear. the model got cheap, taste didn't. i've seen way too many "ai films" with flawless skin and zero blocking
@AdityaKrishnanP@percyliang the recipe framing is clean but it hides the hardest part imo. compute tells you batch size and lr, it does not tell you your data mix is garbage. that's usually where runs actually die
Anthropic's $80.7B net loss is mostly an accounting artifact — $72.4B of it is non-cash marks on convertible notes, and it grows precisely because the company keeps getting more valuable. The number that actually constrains the business is $265.6B.
That's total compute and infrastructure commitments as of June 30, 2026, per Note 13: "approximately $265.6 billion with terms ranging from one to 20 years." At year-end 2025 the same line was $54.6B. Roughly 5x in six months.
It reads like a bill. It isn't. The mechanics differ contract to contract:
Google ≥$114.5B (to 2033), Amazon ≥$110B (to 2036), Microsoft ≥$44.9B (to 2033) — all "regardless of actual usage," underspend settled as a cash true-up at the end, not prepaid. The most rigid part of the stack.
SpaceXAI up to $120.5B, ~$3.1B/month once capacity lands (expected Dec 2026). Non-cancelable only until Jan/Mar 2027, then either side can walk on 90 days' notice.
Nscale ($44.6B) and Lambda ($41.3B) accrue only as capacity is delivered and accepted. Pay-as-you-go wearing a commitment's clothes.
Broadcom is the one that behaves like debt: $161.2B of equipment leases, rent quarterly, and the meter only starts when hardware is "ready for service." Separately, $321.4B of leases signed but not commenced, up from $50.2B a year earlier.
So the question isn't whether they can pay. It's what the bill looks like in 2028-2030, when the ramp years stack.
Which is exactly what the filing won't say. Both figures you'd need are redacted: "approximately $[ ] expected to be incurred before December 31, 2030" and "approximately $[ ] ... due in the next 12 months."
Scale check: H1 2026 compute spend was $8.9B, ~$18B annualized. Spread $265.6B over 10-20 years, it's $13-26B/yr. Same order as today. Not a cliff — a decade locked in on the bet that revenue outruns it.
And the number is already stale: since June 30 they've signed ~$245B more in minimum obligations. None of it is in the $265.6B everyone is quoting.
Anthropic has filed an S-1 registration statement with the SEC related to its proposed IPO.
Read more: https://t.co/ofhv0DK3Zp
Legal disclaimer: A registration statement relating to these securities has been filed with the SEC but has not yet become effective. These securities may not be sold, nor may offers to buy be accepted, prior to the time the registration statement becomes effective. This post does not constitute an offer to sell or the solicitation of an offer to buy these securities, nor shall there be any sale of these securities in any state or jurisdiction in which such offer, solicitation or sale would be unlawful prior to registration or qualification under the securities laws of any such state or jurisdiction.
Kimi K3.1 Pro is the best open-weight model in the world. The weights do not exist yet.
Both of those statements are true on October 24, 2026. Moonshot AI published its benchmarks, opened API access at $2/$10 per million tokens, and promised full weights by November 4. So every "best open model ever" headline you have read this week is describing a download link that is not live. Hugging Face's moonshotai organization still lists only Kimi-K3. The license text has not been published.
This matters less than the usual caveat-emptor posturing, because the numbers behind it are unusually well corroborated. But it matters, and it should sit at the top.
WHAT ACTUALLY SHIPPED
Kimi K3.1 Pro is not a new architecture. Moonshot says it plainly: "a post-trained version of Kimi K3, with substantially larger post-training compute." The base is unchanged. 2.8 trillion total parameters, 104 billion active, 896 experts with 16 firing per token, Kimi Delta Attention plus Attention Residuals, a 1,048,576-token context window, text and image in.
The only new thing is reinforcement learning, run jointly across what the company calls "hundreds of downstream domains." Moonshot will not say how many tokens or GPUs that took. What it will describe is infrastructure: the RL framework, long-context inference, training efficiency, sandbox provisioning, rollout scheduling, and a shared platform for checking environment quality. The claim is that each unit of RL compute now buys a predictable return rather than a noisy one.
There is also a Highspeed variant: same model, 2-3x faster, twice the price on input and output, identical price on cache hits.
THE GAIN IS NOT WHERE YOU EXPECT IT
If K3.1 were simply "more post-training," you would expect coding to improve. It already was competitive. The actual jump is in agentic work — the ability to carry a task across many steps and finish it.
GDPval-AA: 1,534 to 1,846 Elo, a 312-point move that takes it from mid-pack to third overall, behind only Claude Opus 5.5 (1,866) and Muse Wonder Full (1,857).
AutomationBench: 58% to 73%.
That is the difference between a model that writes code and one that closes a ticket. It is also the single hardest thing to buy with more parameters, and the most valuable thing if you are building agents.
On Terminal-Bench 4.0, Moonshot reports 65.7% against Claude Opus 5.5 at 66.4% and GPT-6 Astra at 58.2%. Behind, but inside a percentage point.
Then Artificial Analysis ran the same benchmark independently and got 67% — the highest score it has recorded for any model, including the proprietary frontier.
Two organizations, two harnesses, two different numbers, one conclusion. That agreement is rarer than the score.
On the AA Intelligence Index v4.3.2, K3.1 Pro lands at 55, fourth of 215 models, up 11 points from K3's 44. Above it: Opus 5.5 at 58, Sonnet 5.5 and Muse Wonder Full at 56. Below it, all tied at 53: GPT-6 Astra, Gemini 4 Argon, Claude Fable 5.1.
Read that ordering again. A model whose weights you cannot download is sitting three points under the best proprietary system in the world and above OpenAI's and Google's flagships.
WHAT IT STILL CANNOT DO
The release blog has no limitations section. K3's did. That absence is worth noticing, because the weaknesses are measurable and they are specific.
Terminal-Bench Science 0.1: 38.1%, against GPT-6 Astra at 68.1% and Opus 5.5 at 63.3%. That is not a close race. It is a different tier.
HLE-Diamond: 39.4 against Opus's 60.9. ProgramBench binary score: 31.5 against Opus's 65.0.
And on factual reliability, AA's Omniscience Index puts K3.1 at 30 against Opus's 46. Its hallucination rate did fall from 53% to 44%. Falling by nine points is progress. Sitting at 44% is still a system you cannot point at a spreadsheet unsupervised.
The pattern: exceptional at doing, mediocre at knowing. It can drive a terminal and build a scene. Ask it to reason carefully through a hard science problem and it degrades toward the middle of the pack.
PRICING IS THE ACTUAL WEAPON
$2 input, $10 output, $0.20 on cache hits. That is 33% below Kimi K3 across all three rates, half of Claude Opus 5.5, and identical to GPT-6.1 Sol.
Critically, there is no long-context surcharge anywhere inside the million-token window. GPT-6 Astra doubles its entire bill above 272K tokens. For anyone doing repository-scale or document-scale work, that one line is worth more than the benchmark table.
But cheap per token is not cheap per task. Artificial Analysis measured roughly 150 million output tokens to complete its full index run, at about 81 tokens per second. Cost per task came out at $1.98 — essentially the same as K3, because the lower token price was cancelled by the model burning more tokens. It also means it is slow, and always in thinking mode: requests with thinking disabled are rejected outright.
You are not buying a cheaper model. You are buying a model that thinks longer for the same money.
THE STRANGEST SECTION IN THE RELEASE
Buried past the finance benchmarks is a section on cybersecurity, and it is the most honest thing Moonshot has published.
The company found that offensive security capability was rising steadily as post-training scaled. They traced the source: cyber content embedded in the general mid-training corpus, and cyber tasks present in the RL environment mix. Then they ran the experiment. A checkpoint trained without any mitigation solves 90.5% of their internal xdev-bench-hard tasks at pass@3. After removing cyber data from both stages, the shipped model solves 81.0%.
They deliberately shipped a weaker model, measured exactly how much weaker, and wrote the number down.
The stated principle is that they will not release open weights that extend the absolute frontier of offensive cyber capability, and that they intend to raise that ceiling gradually over time. Whether you find that reassuring or paternalistic probably depends on what you do for a living.
THE DETAIL THAT EXPLAINS THE MODEL
The best thing in the entire post is not a benchmark.
Moonshot took human-atlas, an open anatomical model with 296 skeletal components including cartilage and 402 muscle components, and had K3.1 build a walking controller on top of it.
The simulated human started by collapsing within seconds. Over 167 missions and 420 independent reviews, the model analyzed results, diagnosed failures, and iterated. It ended up walking continuously with alternating strides for 11.2 seconds.
Here is the part that matters. When reward shaping failed to fix the arm motion, K3.1 did not keep tuning the reward. It ran open-loop probes of the simulator's dynamics, found 36 errors in the shoulder-muscle moment-arm calculations, corrected them, and retrained the controller.
That is a different capability from writing code. Writing code is producing a thing that matches a spec. This is looking at a system that failed, suspecting the model of the world is wrong rather than the policy, and going and checking. It is the thing junior engineers take years to learn.
WHAT USERS ARE ACTUALLY COMPLAINING ABOUT
The complaints on Hacker News and in the Chinese developer forums are almost never about output quality. They are about economics and manners.
A $99 subscriber reported burning a five-hour quota in 47 minutes in a single session. Another said the HighSpeed variant does not last two hours on a five-hour allowance. One task consumed 73 tool calls where GPT-6.1 Sol needed 18. And several reported the model editing files without being asked — a known trait inherited from K3, which Moonshot's own K3 blog flagged as "excessive proactiveness."
Great model. Expensive habits.
WHERE THIS ACTUALLY LANDS
On the Intelligence Index, the gap between the best open-weight model and the best proprietary one was 13 points a year ago, 6 points in April 2026, and is 3 points the moment these weights ship.
Mozilla's September analysis put the best open model about 4.4 months behind the closed frontier at roughly 30% of the price. UK AISI measured the cyber-capability gap at 4-7 months, down from 6-10 months through most of 2025.
Those are the same finding from three directions: the open frontier is now roughly a fiscal quarter behind, not a generation.
Which is why the interesting question is no longer "can open models compete." It is what happens to pricing when the second-best model in the world is free to download, and to the business models built on the assumption that it never would be.
One caveat on all of the above. Terminal-Bench 4.0 has been audited as close to saturated — 41 of its 65 tasks reportedly fail even when solved honestly. Sonnet 5.5 moved from 62.8% to 89.2% purely by changing harness. So treat absolute scores as directional. The rankings here survive that skepticism better than most, because two independent parties landed in the same place.
Kimi K3.1 Pro is the first open model where "just use the closed one" stopped being the obvious default.
Assuming the weights actually ship.
only one of the three has a clock attached. a reactor is 10 years from signature to first watt and most of that is permits, transmission queues and interconnection studies, not concrete. if the goal is vitality before the next winter you build interconnects and storage, not a reactor. the other two steps cost a press release.