Watch the full webinar (46 min): the million-token context window, parallel agents, the fast mode to skip, and where to point this at construction admin work.
The honest summary of Opus 4.8: it is a marginal improvement over the version before it, and a few of those margins land squarely on construction documentation.
That is worth knowing either way, because the model releases from labs like Claude by Anthropic are going to keep coming in an arms race between the major labs, and evaluating each one on what it changes for your actual work is more useful than reading benchmark headlines. The next release in that cadence is covered in Claude Fable 5 for Construction, the other side of the race in Which ChatGPT Model to Use for Construction Work, and the field overall in AI for Construction: The Complete Guide and Best AI for Construction: A Comparison.
Here is what changed, what to use it for, and the one new setting to leave alone.
The Million-Token Context Window
This is the headline feature, and understanding it requires understanding the problem it addresses.
AI has a short-term memory problem. Every time it answers, it has to reference the full conversation history and every document you have given it. Chat long enough, or upload enough, and it can no longer hold all of it. When it cannot hold everything, it starts filling gaps. That is what hallucination is.
A token is roughly a character, so a million-token context window is easiest to picture as about a million characters of working memory. That does not eliminate the memory problem. It extends it substantially: you can now remember for three days instead of one, to use the rough analogy.
The frustrating part is that you cannot easily see when you are approaching the limit in a normal chat session. The best signal available is when the system tells you it is compressing and summarizing the conversation, and a summary is exactly where a critical project detail can quietly disappear. Watch for it.
What This Unlocks for Contract Review
The practical construction win: an entire contract, including all exhibits and every change order, held in memory at once.
That eliminates chunking for documents in that size range. Chunking is the standard workaround where a long document is split into pieces, stored, and retrieved as needed, because it is expensive to re-read a large PDF on every question. Chunking works. It is also where nuance gets lost, because the model sees a slice rather than the whole agreement.
With the full document in a single pass, you can compare the original contract against every change order at once and ask the questions that matter: where did liability shift, where did insurance requirements change, where did payment terms move. Those answers are much more reliable when the model is looking at the entire document rather than reasoning across retrieved fragments.
One clarification, because it gets overstated. A million tokens is not your document archive. A firm with 100,000 issued documents still needs a chunked, indexed store to answer questions across all of them. The context window changed single-document review. It did not change the architecture of a knowledge base.
Parallel Agents and Agent Teams
The bigger structural change in this release is dynamic workflows and the ability to run hundreds of agents in parallel. The prior version might spin up around three.
Both existed before. This release put them on steroids.
Here is why it matters. When you ask a complex question, the system often does not answer it directly. It dispatches sub-agents, each researching a piece, each with its own full context window, and a coordinating agent assembles the result. More agents means faster and, importantly, more accurate work, because no single agent is trying to fit everything into one context.
There are two distinct patterns and they are worth telling apart.
Parallel agents work independently. They do not talk to each other. They go, they do their piece, they report back.
Agent teams collaborate. They message each other, share findings mid-task, and adjust based on what the others discover, all under a coordinating agent.
Agent teams are genuinely impressive and they will consume a week’s worth of usage in about twenty minutes. Unless you are building something unusually complex, parallel agents are what you actually need.
Fewer Missed Flaws
The quieter improvement worth knowing: the model is roughly four times less likely to let a flaw slip through compared to the previous version.
That is self-catching, not raw intelligence, and self-catching is what matters for compliance work. Reviewing a certificate of insurance against company standards, updating a workbook, checking a certified payroll: these are tasks where the failure mode is a missed detail rather than a wrong conclusion.
Skip the Fast Mode
The release also added a fast mode delivering around 2.5 times faster responses.
Leave it off for anything that matters. Standard response time is not the bottleneck in a construction back office. Waiting an extra few seconds for a meaningfully better answer is a trade worth making every time.
The parallel is one every builder knows. Done fast and done right are not the same job. Fast mode buys speed by cutting the thinking, and the thinking is what you are paying for.
Same logic applies to the adaptive thinking setting, which lets you dial reasoning effort between low, medium, and high. Leaving it on high is the simpler policy. Choosing low to save a little and then spending ten minutes fixing a mediocre answer is a bad trade.
Accuracy Is Usually a Context Problem, Not a Model Problem
Adoption surveys consistently show a majority of firms citing lack of trust in AI output accuracy as their primary blocker. Meanwhile, roughly 45 percent of construction AI investment is being pointed at office and administrative work, which is where the volume is.
In practice, the accuracy complaints almost always trace back to insufficient context rather than to model capability. Someone asks the AI to respond to an RFI without loading the specification section, the submittal history, the relevant contract language, or the project’s prior correspondence, then concludes the tool is unreliable.
It needed the project. It was given a question.
The counterintuitive implication: the way to raise trust in AI output is to spend your effort on preloading context, not on chasing the newest model.
Where to Point It in a Construction Back Office
Certified payroll. On large prevailing wage jobs governed by Davis-Bacon and Related Acts it is not unusual to find a dedicated full-time employee in the project trailer reviewing certified payroll submitted by vendors. It is among the most labor-intensive manual tasks in this industry. The workable pattern is AI first pass, human decision: the system does all the analysis and returns a green, yellow, or red result. Green means nothing found. Yellow means something worth a look. Red means do not approve without a deliberate exception. A person makes the call in seconds instead of hours. More on the tooling in Certified Payroll Software for Construction.
Submittals. Cross-check against specifications and flag missing information.
Vendor onboarding. Screen an entire prequalification package quickly, verifying compliance and documentation against your requirements and your documented exceptions.
Certificates of insurance. Same green, yellow, red pattern against the ACORD 25 and endorsements, walked through step by step in How AI Reads a Certificate of Insurance.
Notice what these have in common. The AI does the heavy lifting and surfaces a recommendation. The human decides. That is the configuration that gets adopted, because it lets a firm deploy AI without betting the business on it.
How to Roll It Out
Permissions come in three levels and the names are self-explanatory: read-only, ask for approval before each action, and full allow. Start in read-only where the agent has no reason to act. Move to approval-required where it does. Treat full allow as something earned.
Then take it one level at a time. The fastest way to fail with AI is trying to do too much too soon. Firms that succeed crawl, then walk, then run, and they do it in that order because construction teams have been doing the work the same way for years and switching everything mid-project is not realistic.
Start with one bounded job. Certificate of insurance tracking is a good one: high volume, clear rules, low risk, and it proves value quickly enough to build internal trust for the next thing.
You do not need to be a developer for any of it. Most first-day applications are drag and drop with plain English instructions.
Frequently Asked Questions
What does a 1 million token context window mean for construction documents?
A token is roughly a character, so it is about a million characters of working memory. Practically, it means an entire contract plus its exhibits and every change order can be held at once, so the model reads the whole agreement rather than retrieved fragments. That makes questions about liability shifts, insurance requirement changes, and payment term adjustments far more reliable. It does not mean the model can hold your document archive; a large library still needs an indexed store.
Why does AI hallucinate on construction projects?
Because it runs out of room. Every response requires referencing the full conversation and every document supplied. Once that exceeds the context window, the model can no longer hold all of it and begins filling gaps with plausible content. Longer context windows reduce how often this happens. The other common cause is simply not being given the project context needed to answer correctly in the first place.
What is the difference between parallel agents and agent teams?
Parallel agents work independently on separate pieces of a task, each with its own context, and report back to a coordinating agent without talking to each other. Agent teams collaborate: they message each other, share findings mid-task, and adjust based on what the others discover. Teams are more capable and consume dramatically more usage. For most construction back-office work, parallel agents are sufficient.
Should I use fast mode?
Not for work that matters. Fast mode delivers roughly 2.5 times faster responses by cutting reasoning, and response speed is rarely the bottleneck in a back office. Waiting a few extra seconds for a better answer is almost always the right trade. The same reasoning applies to the low, medium, and high thinking settings: leaving it on high avoids the loop of a weak answer followed by ten minutes of correction.
How do I keep control over what an AI agent can do?
Three permission levels. Read-only, where it can look and report but not act. Approval required, where it asks before every action. And full allow, where it acts without asking. Start in read-only anywhere the agent has no reason to change anything, use approval-required where it does, and treat full allow as something earned over time on low-risk, high-repetition work.
Green, Yellow, Red on Every Compliance Document
Contracts Connected runs the first pass on your certificates of insurance and certified payroll, checks them against your project requirements and the endorsement language, and hands your team a green, yellow, or red result with the reasoning attached. The decision stays with a person.
No pitch deck. Straight conversation about what you’re managing and whether we can help.